Sygil-Dev/sygil-webui
1
1
Fork 0
mirror of https://github.com/Sygil-Dev/sygil-webui.git synced 2025-01-05 19:04:06 +03:00
Code Issues Projects Releases Wiki Activity

Repo merge (#712)

Browse Source 
* repo-merge

* cutdown size

* Create setup.py

* webui.cmd

* ldm

* Update environment.yaml

* Update environment.yaml
This commit is contained in:
hlky 2022-09-06 23:50:14 +01:00 committed by GitHub
parent 78ad3c3445
commit 010b27ce9a
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
94 changed files with 16478 additions and 4574 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
.gitattributes vendored Normal file
View File

@ -0,0 +1,4 @@
* text=auto
*.{cmd,[cC][mM][dD]} text eol=crlf
*.{bat,[bB][aA][tT]} text eol=crlf
*.sh text eol=lf

29
.github/sync.yml vendored
View File

@ -1,29 +0,0 @@
hlky/stable-diffusion:
- source: webui.py
dest: scripts/webui.py
- source: relauncher.py
dest: scripts/relauncher.py
- source: txt2img.yaml
dest: txt2img.yaml
- source: webui.yaml
dest: configs/webui/webui.yaml
- source: readme.md
dest: readme.md
- source: .gitignore/.idea/.gitignore
dest: .gitignore/.idea/.gitignore
- source: frontend/__init__.py
dest: frontend/__init__.py
- source: frontend/css_and_js.py
dest: frontend/css_and_js.py
- source: frontend/frontend.py
dest: frontend/frontend.py
- source: frontend/ui_functions.py
dest: frontend/ui_functions.py
- source: frontend/job_manager.py
dest: frontend/job_manager.py
- source: frontend/css/no_progress_bar.css
dest: frontend/css/no_progress_bar.css
- source: frontend/css/styles.css
dest: frontend/css/styles.css
- source: frontend/js/index.js
dest: frontend/js/index.js

16
.github/workflows/sync.yml vendored
View File

@ -1,16 +0,0 @@
name: Sync Files
on:
push:
branches:
- master
workflow_dispatch:
jobs:
sync:
runs-on: ubuntu-latest
steps:
- name: Checkout Repository
uses: actions/checkout@master
- name: Run GitHub File Sync
uses: BetaHuhn/repo-file-sync-action@v1
with:
GH_PAT: ${{ secrets.GH_PAT }}

7
.gitignore vendored
View File

@ -43,6 +43,7 @@ share/python-wheels/
.installed.cfg
*.egg
MANIFEST
.env_docker
.env_updated
condaenv.*.requirements.txt
@ -50,3 +51,9 @@ condaenv.*.requirements.txt
# =========================================================================== #
# Repo-specific
# =========================================================================== #
/custom-conda-path.txt
/src/*
/outputs/*
/log/**/*.png
/log/log.csv
/flagged/*

26
Dockerfile Normal file
View File

@ -0,0 +1,26 @@
FROM nvidia/cuda:11.3.1-runtime-ubuntu20.04
ENV DEBIAN_FRONTEND=noninteractive
WORKDIR /sd
SHELL ["/bin/bash", "-c"]
RUN apt-get update && \
apt-get install -y libglib2.0-0 wget && \
apt-get clean && \
rm -rf /var/lib/apt/lists/*
# Install miniconda
ENV CONDA_DIR /opt/conda
RUN wget -O ~/miniconda.sh -q --show-progress --progress=bar:force https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh && \
/bin/bash ~/miniconda.sh -b -p $CONDA_DIR && \
rm ~/miniconda.sh
ENV PATH=$CONDA_DIR/bin:$PATH
# Install font for prompt matrix
COPY /data/DejaVuSans.ttf /usr/share/fonts/truetype/
EXPOSE 7860
COPY ./entrypoint.sh /sd/
ENTRYPOINT /sd/entrypoint.sh

175
README.md
View File

@ -1,29 +1,40 @@
### [MAIN REPO](https://github.com/hlky/stable-diffusion)
[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/altryne/sd-webui-colab/blob/main/Stable_Diffusion_WebUi_Altryne.ipynb)
# This repo is for development, there may be bugs and new features
# [Installation](https://github.com/hlky/stable-diffusion/wiki/Installation)
# Notice
-New LDSR settings added to Image Lab, To use the new LDSR settings please make sure to re-clone the LDSR (Instructions added below) to insure you have the latest.
## [Development, testing, bleeding edge, maybe have bugs](https://github.com/hlky/stable-diffusion-webui)
## Feature request? Use [discussions](https://github.com/hlky/stable-diffusion-webui/discussions)
### Have an **issue**?
* If the issue involves _a bug_ in **textual-inversion** create the issue on **_[hlky/stable-diffusion-webui](https://github.com/hlky/stable-diffusion-webui)_**
* If you want to know how to **activate** or **use** textual-inversion see **_[hlky/sd-enable-textual-inversion](https://github.com/hlky/sd-enable-textual-inversion)_**. Activation not working? create the issue on **_[hlky/stable-diffusion-webui](https://github.com/hlky/stable-diffusion-webui)_**
## More documentation about features, troubleshooting, common issues very soon
### Want to help with documentation? Documented something? Use [Discussions](https://github.com/hlky/stable-diffusion/discussions)
## **Important**
🔥 NEW! webui.cmd updates with any changes in environment.yaml file so the environment will always be up to date as long as you get the new environment.yaml file 🔥
:fire: no need to remove environment, delete src folder and create again, MUCH simpler! 🔥
--------------
### Questions about **_[Upscalers](https://github.com/hlky/stable-diffusion-webui/wiki/Upscalers)_**?
### Questions about **_[Optimized mode](https://github.com/hlky/stable-diffusion-webui/wiki/Optimized-mode)_**?
### Questions about **_[Command line options](https://github.com/hlky/stable-diffusion-webui/wiki/Command-line-options)_**?
--------------
[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/altryne/sd-webui-colab/blob/main/Stable_Diffusion_WebUi_Altryne.ipynb)
## More documentation about features, troubleshooting, common issues very soon
### Want to help with documentation? Documented something? Use [Discussions](https://github.com/hlky/stable-diffusion-webui/discussions)
Features:
* Gradio GUI: Idiot-proof, fully featured frontend for both txt2img and img2img generation
* No more manually typing parameters, now all you have to do is write your prompt and adjust sliders
* :fire: :fire: Optimized support!! :fire: :fire:
* 🔥 NEW! [webui.cmd](https://github.com/hlky/stable-diffusion) updates with any changes in environment.yaml file so the environment will always be up to date as long as you get the new environment.yaml file 🔥
:fire: no need to remove environment, delete src folder and create again, MUCH simpler! 🔥
* GFPGAN Face Correction 🔥: [Download the model](https://github.com/hlky/stable-diffusion-webui#gfpgan)Automatically correct distorted faces with a built-in GFPGAN option, fixes them in less than half a second
* RealESRGAN Upscaling 🔥: [Download the models](https://github.com/hlky/stable-diffusion-webui#realesrgan) Boosts the resolution of images with a built-in RealESRGAN option
* :computer: esrgan/gfpgan on cpu support :computer:
@ -47,13 +58,12 @@ Features:
# Stable Diffusion web UI
A browser interface based on Gradio library for Stable Diffusion.
![](images/txt2img.jpg)
Original script with Gradio UI was written by a kind anonymous user. This is a modification.
![](images/img2img.jpg)
![](images/gfpgan.jpg)
![](images/esrgan.jpg)
![](https://github.com/hlky/stable-diffusion-webui/blob/master/images/txt2img.jpg)
![](https://github.com/hlky/stable-diffusion-webui/blob/master/images/img2img.jpg)
![](https://github.com/hlky/stable-diffusion-webui/blob/master/images/gfpgan.jpg)
![](https://github.com/hlky/stable-diffusion-webui/blob/master/images/esrgan.jpg)
### GFPGAN
@ -65,15 +75,6 @@ into the `/stable-diffusion/src/gfpgan/experiments/pretrained_models` directory.
Download [RealESRGAN_x4plus.pth](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth) and [RealESRGAN_x4plus_anime_6B.pth](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth).
Put them into the `stable-diffusion/src/realesrgan/experiments/pretrained_models` directory.
### LDSR
Quadruple your resolution using Latent Diffusion, to install:
- Git clone https://github.com/devilismyfriend/latent-diffusion into your stable-diffusion-main/src/ folder
- Rename latent-diffusion-main folder to latent-diffusion
- If on windows: run download_models.bat to download the required model files
- Otherwise to manually install the model download [project.yaml](https://heibox.uni-heidelberg.de/f/31a76b13ea27482981b4/?dl=1) and [last.cpkt](https://heibox.uni-heidelberg.de/f/578df07c8fc04ffbadf3/?dl=1) and rename last.ckpt to model.ckpt
- Place both under stable-diffusion-main/src/latent-diffusion/experiments/pretrained_models/
- Make sure you have both project.yaml and model.ckpt in that folder and path.
- LDSR should be wokring now.
### Web UI
When launching, you may get a very long warning message related to some weights not being used. You may freely ignore it.
@ -99,7 +100,6 @@ also a separate tab that just allows you to use GFPGAN on any picture, with a sl
Lets you double the resolution of generated images. There is a checkbox in every tab to use RealESRGAN, and you can choose between the regular upscaler and the anime version.
There is also a separate tab for using RealESRGAN on any picture.
![](images/RealESRGAN.png)
### Sampling method selection
@ -123,19 +123,6 @@ Four images will be produced, in this order, all with same seed and each with co
Another example, this time with 5 prompts and 16 variations:
![](images/prompt_matrix.jpg)
### Prompt combinations
If you add '@' symbol at start your prompt and change text like this:
`@(moba|rpg|rts) character (2d|3d) model` it will be produce 3 * 2 combinations or prompt with same seed:
- `moba character 2d model`
- `rpg character 2d model`
- `rts character 2d model`
- `moba character 3d model`
- `rpg character 3d model`
- `rts character 3d model`
If you use this feature, batch count will be ignored, because the number of pictures to produce
depends on your prompts, but batch size will still work (generating multiple pictures at the
same time for a small speed boost).
@ -188,18 +175,108 @@ saving output image, and replacing input image with it. Batch count setting cont
this you get.
Usually, when doing this, you would choose one of many images for the next iteration yourself, so the usefulness
of this feature may be questionable, but I've managed to get some very nice outputs with it that I wasn't able
of this feature may be questionable, but I've managed to get some very nice outputs with it that I wasn't abble
to get otherwise.
Example: (cherrypicked result; original picture by anon)
![](images/loopback.jpg)
## Development Info
There is a different directory structure on this dev repo to simplify things and a github action is used to sync things to the right place in the [main repo](https://github.com/hlky/stable-diffusion). The config for this sync is in `.github/sync.yml`.
There is a helper script for local development to replicate the actions of this github action.
- Run `python sync_local.py --dest MAIN_REPO_FOLDER` to replicate the effect of this sync.
- To copy changes back you can run `python sync_local.py --dest MAIN_REPO_FOLDER --reverse`.
### --help
```
optional arguments:
-h, --help show this help message and exit
--outdir [OUTDIR] dir to write results to
--outdir_txt2img [OUTDIR_TXT2IMG]
dir to write txt2img results to (overrides --outdir)
--outdir_img2img [OUTDIR_IMG2IMG]
dir to write img2img results to (overrides --outdir)
--save-metadata Whether to embed the generation parameters in the sample images
--skip-grid do not save a grid, only individual samples. Helpful when evaluating lots of samples
--skip-save do not save indiviual samples. For speed measurements.
--n_rows N_ROWS rows in the grid; use -1 for autodetect and 0 for n_rows to be same as batch_size (default:
-1)
--config CONFIG path to config which constructs model
--ckpt CKPT path to checkpoint of model
--precision {full,autocast}
evaluate at this precision
--gfpgan-dir GFPGAN_DIR
GFPGAN directory
--realesrgan-dir REALESRGAN_DIR
RealESRGAN directory
--realesrgan-model REALESRGAN_MODEL
Upscaling model for RealESRGAN
--no-verify-input do not verify input to check if it's too long
--no-half do not switch the model to 16-bit floats
--no-progressbar-hiding
do not hide progressbar in gradio UI (we hide it because it slows down ML if you have hardware
accleration in browser)
--defaults DEFAULTS path to configuration file providing UI defaults, uses same format as cli parameter
--gpu GPU choose which GPU to use if you have multiple
--extra-models-cpu run extra models (GFGPAN/ESRGAN) on cpu
--esrgan-cpu run ESRGAN on cpu
--gfpgan-cpu run GFPGAN on cpu
--cli CLI don't launch web server, take Python function kwargs from this file.
```
-----
# Stable Diffusion
*Stable Diffusion was made possible thanks to a collaboration with [Stability AI](https://stability.ai/) and [Runway](https://runwayml.com/) and builds upon our previous work:*
[**High-Resolution Image Synthesis with Latent Diffusion Models**](https://ommer-lab.com/research/latent-diffusion-models/)<br/>
[Robin Rombach](https://github.com/rromb)\*,
[Andreas Blattmann](https://github.com/ablattmann)\*,
[Dominik Lorenz](https://github.com/qp-qp)\,
[Patrick Esser](https://github.com/pesser),
[Björn Ommer](https://hci.iwr.uni-heidelberg.de/Staff/bommer)<br/>
**CVPR '22 Oral**
which is available on [GitHub](https://github.com/CompVis/latent-diffusion). PDF at [arXiv](https://arxiv.org/abs/2112.10752). Please also visit our [Project page](https://ommer-lab.com/research/latent-diffusion-models/).
![txt2img-stable2](assets/stable-samples/txt2img/merged-0006.png)
[Stable Diffusion](#stable-diffusion-v1) is a latent text-to-image diffusion
model.
Thanks to a generous compute donation from [Stability AI](https://stability.ai/) and support from [LAION](https://laion.ai/), we were able to train a Latent Diffusion Model on 512x512 images from a subset of the [LAION-5B](https://laion.ai/blog/laion-5b/) database.
Similar to Google's [Imagen](https://arxiv.org/abs/2205.11487),
this model uses a frozen CLIP ViT-L/14 text encoder to condition the model on text prompts.
With its 860M UNet and 123M text encoder, the model is relatively lightweight and runs on a GPU with at least 10GB VRAM.
See [this section](#stable-diffusion-v1) below and the [model card](https://huggingface.co/CompVis/stable-diffusion).
## Stable Diffusion v1
Stable Diffusion v1 refers to a specific configuration of the model
architecture that uses a downsampling-factor 8 autoencoder with an 860M UNet
and CLIP ViT-L/14 text encoder for the diffusion model. The model was pretrained on 256x256 images and
then finetuned on 512x512 images.
*Note: Stable Diffusion v1 is a general text-to-image diffusion model and therefore mirrors biases and (mis-)conceptions that are present
in its training data.
Details on the training procedure and data, as well as the intended use of the model can be found in the corresponding [model card](https://huggingface.co/CompVis/stable-diffusion).
## Comments
- Our codebase for the diffusion models builds heavily on [OpenAI's ADM codebase](https://github.com/openai/guided-diffusion)
and [https://github.com/lucidrains/denoising-diffusion-pytorch](https://github.com/lucidrains/denoising-diffusion-pytorch).
Thanks for open-sourcing!
- The implementation of the transformer encoder is from [x-transformers](https://github.com/lucidrains/x-transformers) by [lucidrains](https://github.com/lucidrains?tab=repositories).
## BibTeX
```
@misc{rombach2021highresolution,
title={High-Resolution Image Synthesis with Latent Diffusion Models},
author={Robin Rombach and Andreas Blattmann and Dominik Lorenz and Patrick Esser and Björn Ommer},
year={2021},
eprint={2112.10752},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
You can the file `webui_playground.py`, which does not load the models, to more rapidly iterate on UI changes and then copy those changes into `webui.py`,

140
Stable_Diffusion_v1_Model_Card.md Normal file
View File

@ -0,0 +1,140 @@
# Stable Diffusion v1 Model Card
This model card focuses on the model associated with the Stable Diffusion model, available [here](https://github.com/CompVis/stable-diffusion).
## Model Details
- **Developed by:** Robin Rombach, Patrick Esser
- **Model type:** Diffusion-based text-to-image generation model
- **Language(s):** English
- **License:** [Proprietary](LICENSE)
- **Model Description:** This is a model that can be used to generate and modify images based on text prompts. It is a [Latent Diffusion Model](https://arxiv.org/abs/2112.10752) that uses a fixed, pretrained text encoder ([CLIP ViT-L/14](https://arxiv.org/abs/2103.00020)) as suggested in the [Imagen paper](https://arxiv.org/abs/2205.11487).
- **Resources for more information:** [GitHub Repository](https://github.com/CompVis/stable-diffusion), [Paper](https://arxiv.org/abs/2112.10752).
- **Cite as:**
@InProceedings{Rombach_2022_CVPR,
author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},
title = {High-Resolution Image Synthesis With Latent Diffusion Models},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2022},
pages = {10684-10695}
}
# Uses
## Direct Use
The model is intended for research purposes only. Possible research areas and
tasks include
- Safe deployment of models which have the potential to generate harmful content.
- Probing and understanding the limitations and biases of generative models.
- Generation of artworks and use in design and other artistic processes.
- Applications in educational or creative tools.
- Research on generative models.
Excluded uses are described below.
### Misuse, Malicious Use, and Out-of-Scope Use
_Note: This section is taken from the [DALLE-MINI model card](https://huggingface.co/dalle-mini/dalle-mini), but applies in the same way to Stable Diffusion v1_.
The model should not be used to intentionally create or disseminate images that create hostile or alienating environments for people. This includes generating images that people would foreseeably find disturbing, distressing, or offensive; or content that propagates historical or current stereotypes.
#### Out-of-Scope Use
The model was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model.
#### Misuse and Malicious Use
Using the model to generate content that is cruel to individuals is a misuse of this model. This includes, but is not limited to:
- Generating demeaning, dehumanizing, or otherwise harmful representations of people or their environments, cultures, religions, etc.
- Intentionally promoting or propagating discriminatory content or harmful stereotypes.
- Impersonating individuals without their consent.
- Sexual content without consent of the people who might see it.
- Mis- and disinformation
- Representations of egregious violence and gore
- Sharing of copyrighted or licensed material in violation of its terms of use.
- Sharing content that is an alteration of copyrighted or licensed material in violation of its terms of use.
## Limitations and Bias
### Limitations
- The model does not achieve perfect photorealism
- The model cannot render legible text
- The model does not perform well on more difficult tasks which involve compositionality, such as rendering an image corresponding to “A red cube on top of a blue sphere”
- Faces and people in general may not be generated properly.
- The model was trained mainly with English captions and will not work as well in other languages.
- The autoencoding part of the model is lossy
- The model was trained on a large-scale dataset
[LAION-5B](https://laion.ai/blog/laion-5b/) which contains adult material
and is not fit for product use without additional safety mechanisms and
considerations.
### Bias
While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases.
Stable Diffusion v1 was trained on subsets of [LAION-2B(en)](https://laion.ai/blog/laion-5b/),
which consists of images that are primarily limited to English descriptions.
Texts and images from communities and cultures that use other languages are likely to be insufficiently accounted for.
This affects the overall output of the model, as white and western cultures are often set as the default. Further, the
ability of the model to generate content with non-English prompts is significantly worse than with English-language prompts.
## Training
**Training Data**
The model developers used the following dataset for training the model:
- LAION-2B (en) and subsets thereof (see next section)
**Training Procedure**
Stable Diffusion v1 is a latent diffusion model which combines an autoencoder with a diffusion model that is trained in the latent space of the autoencoder. During training,
- Images are encoded through an encoder, which turns images into latent representations. The autoencoder uses a relative downsampling factor of 8 and maps images of shape H x W x 3 to latents of shape H/f x W/f x 4
- Text prompts are encoded through a ViT-L/14 text-encoder.
- The non-pooled output of the text encoder is fed into the UNet backbone of the latent diffusion model via cross-attention.
- The loss is a reconstruction objective between the noise that was added to the latent and the prediction made by the UNet.
We currently provide three checkpoints, `sd-v1-1.ckpt`, `sd-v1-2.ckpt` and `sd-v1-3.ckpt`,
which were trained as follows,
- `sd-v1-1.ckpt`: 237k steps at resolution `256x256` on [laion2B-en](https://huggingface.co/datasets/laion/laion2B-en).
194k steps at resolution `512x512` on [laion-high-resolution](https://huggingface.co/datasets/laion/laion-high-resolution) (170M examples from LAION-5B with resolution `>= 1024x1024`).
- `sd-v1-2.ckpt`: Resumed from `sd-v1-1.ckpt`.
515k steps at resolution `512x512` on "laion-improved-aesthetics" (a subset of laion2B-en,
filtered to images with an original size `>= 512x512`, estimated aesthetics score `> 5.0`, and an estimated watermark probability `< 0.5`. The watermark estimate is from the LAION-5B metadata, the aesthetics score is estimated using an [improved aesthetics estimator](https://github.com/christophschuhmann/improved-aesthetic-predictor)).
- `sd-v1-3.ckpt`: Resumed from `sd-v1-2.ckpt`. 195k steps at resolution `512x512` on "laion-improved-aesthetics" and 10\% dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
- **Hardware:** 32 x 8 x A100 GPUs
- **Optimizer:** AdamW
- **Gradient Accumulations**: 2
- **Batch:** 32 x 8 x 2 x 4 = 2048
- **Learning rate:** warmup to 0.0001 for 10,000 steps and then kept constant
## Evaluation Results
Evaluations with different classifier-free guidance scales (1.5, 2.0, 3.0, 4.0,
5.0, 6.0, 7.0, 8.0) and 50 PLMS sampling
steps show the relative improvements of the checkpoints:
![pareto](assets/v1-variants-scores.jpg)
Evaluated using 50 PLMS steps and 10000 random prompts from the COCO2017 validation set, evaluated at 512x512 resolution. Not optimized for FID scores.
## Environmental Impact
**Stable Diffusion v1** **Estimated Emissions**
Based on that information, we estimate the following CO2 emissions using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700). The hardware, runtime, cloud provider, and compute region were utilized to estimate the carbon impact.
- **Hardware Type:** A100 PCIe 40GB
- **Hours used:** 150000
- **Cloud Provider:** AWS
- **Compute Region:** US-east
- **Carbon Emitted (Power consumption x Time x Carbon produced based on location of power grid):** 11250 kg CO2 eq.
## Citation
@InProceedings{Rombach_2022_CVPR,
author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},
title = {High-Resolution Image Synthesis With Latent Diffusion Models},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2022},
pages = {10684-10695}
}
*This model card was written by: Robin Rombach and Patrick Esser and is based on the [DALL-E Mini model card](https://huggingface.co/dalle-mini/dalle-mini).*

54
configs/autoencoder/autoencoder_kl_16x16x16.yaml Normal file
View File

@ -0,0 +1,54 @@
model:
base_learning_rate: 4.5e-6
target: ldm.models.autoencoder.AutoencoderKL
params:
monitor: "val/rec_loss"
embed_dim: 16
lossconfig:
target: ldm.modules.losses.LPIPSWithDiscriminator
params:
disc_start: 50001
kl_weight: 0.000001
disc_weight: 0.5
ddconfig:
double_z: True
z_channels: 16
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult: [ 1,1,2,2,4] # num_down = len(ch_mult)-1
num_res_blocks: 2
attn_resolutions: [16]
dropout: 0.0
data:
target: main.DataModuleFromConfig
params:
batch_size: 12
wrap: True
train:
target: ldm.data.imagenet.ImageNetSRTrain
params:
size: 256
degradation: pil_nearest
validation:
target: ldm.data.imagenet.ImageNetSRValidation
params:
size: 256
degradation: pil_nearest
lightning:
callbacks:
image_logger:
target: main.ImageLogger
params:
batch_frequency: 1000
max_images: 8
increase_log_steps: True
trainer:
benchmark: True
accumulate_grad_batches: 2

53
configs/autoencoder/autoencoder_kl_32x32x4.yaml Normal file
View File

@ -0,0 +1,53 @@
model:
base_learning_rate: 4.5e-6
target: ldm.models.autoencoder.AutoencoderKL
params:
monitor: "val/rec_loss"
embed_dim: 4
lossconfig:
target: ldm.modules.losses.LPIPSWithDiscriminator
params:
disc_start: 50001
kl_weight: 0.000001
disc_weight: 0.5
ddconfig:
double_z: True
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult: [ 1,2,4,4 ] # num_down = len(ch_mult)-1
num_res_blocks: 2
attn_resolutions: [ ]
dropout: 0.0
data:
target: main.DataModuleFromConfig
params:
batch_size: 12
wrap: True
train:
target: ldm.data.imagenet.ImageNetSRTrain
params:
size: 256
degradation: pil_nearest
validation:
target: ldm.data.imagenet.ImageNetSRValidation
params:
size: 256
degradation: pil_nearest
lightning:
callbacks:
image_logger:
target: main.ImageLogger
params:
batch_frequency: 1000
max_images: 8
increase_log_steps: True
trainer:
benchmark: True
accumulate_grad_batches: 2

54
configs/autoencoder/autoencoder_kl_64x64x3.yaml Normal file
View File

@ -0,0 +1,54 @@
model:
base_learning_rate: 4.5e-6
target: ldm.models.autoencoder.AutoencoderKL
params:
monitor: "val/rec_loss"
embed_dim: 3
lossconfig:
target: ldm.modules.losses.LPIPSWithDiscriminator
params:
disc_start: 50001
kl_weight: 0.000001
disc_weight: 0.5
ddconfig:
double_z: True
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult: [ 1,2,4 ] # num_down = len(ch_mult)-1
num_res_blocks: 2
attn_resolutions: [ ]
dropout: 0.0
data:
target: main.DataModuleFromConfig
params:
batch_size: 12
wrap: True
train:
target: ldm.data.imagenet.ImageNetSRTrain
params:
size: 256
degradation: pil_nearest
validation:
target: ldm.data.imagenet.ImageNetSRValidation
params:
size: 256
degradation: pil_nearest
lightning:
callbacks:
image_logger:
target: main.ImageLogger
params:
batch_frequency: 1000
max_images: 8
increase_log_steps: True
trainer:
benchmark: True
accumulate_grad_batches: 2

53
configs/autoencoder/autoencoder_kl_8x8x64.yaml Normal file
View File

@ -0,0 +1,53 @@
model:
base_learning_rate: 4.5e-6
target: ldm.models.autoencoder.AutoencoderKL
params:
monitor: "val/rec_loss"
embed_dim: 64
lossconfig:
target: ldm.modules.losses.LPIPSWithDiscriminator
params:
disc_start: 50001
kl_weight: 0.000001
disc_weight: 0.5
ddconfig:
double_z: True
z_channels: 64
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult: [ 1,1,2,2,4,4] # num_down = len(ch_mult)-1
num_res_blocks: 2
attn_resolutions: [16,8]
dropout: 0.0
data:
target: main.DataModuleFromConfig
params:
batch_size: 12
wrap: True
train:
target: ldm.data.imagenet.ImageNetSRTrain
params:
size: 256
degradation: pil_nearest
validation:
target: ldm.data.imagenet.ImageNetSRValidation
params:
size: 256
degradation: pil_nearest
lightning:
callbacks:
image_logger:
target: main.ImageLogger
params:
batch_frequency: 1000
max_images: 8
increase_log_steps: True
trainer:
benchmark: True
accumulate_grad_batches: 2

86
configs/latent-diffusion/celebahq-ldm-vq-4.yaml Normal file
View File

@ -0,0 +1,86 @@
model:
base_learning_rate: 2.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0195
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: image
image_size: 64
channels: 3
monitor: val/loss_simple_ema
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 64
in_channels: 3
out_channels: 3
model_channels: 224
attention_resolutions:
# note: this isn\t actually the resolution but
# the downsampling factor, i.e. this corresnponds to
# attention on spatial resolution 8,16,32, as the
# spatial reolution of the latents is 64 for f4
- 8
- 4
- 2
num_res_blocks: 2
channel_mult:
- 1
- 2
- 3
- 4
num_head_channels: 32
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 3
n_embed: 8192
ckpt_path: models/first_stage_models/vq-f4/model.ckpt
ddconfig:
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config: __is_unconditional__
data:
target: main.DataModuleFromConfig
params:
batch_size: 48
num_workers: 5
wrap: false
train:
target: taming.data.faceshq.CelebAHQTrain
params:
size: 256
validation:
target: taming.data.faceshq.CelebAHQValidation
params:
size: 256
lightning:
callbacks:
image_logger:
target: main.ImageLogger
params:
batch_frequency: 5000
max_images: 8
increase_log_steps: False
trainer:
benchmark: True

98
configs/latent-diffusion/cin-ldm-vq-f8.yaml Normal file
View File

@ -0,0 +1,98 @@
model:
base_learning_rate: 1.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0195
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: image
cond_stage_key: class_label
image_size: 32
channels: 4
cond_stage_trainable: true
conditioning_key: crossattn
monitor: val/loss_simple_ema
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 32
in_channels: 4
out_channels: 4
model_channels: 256
attention_resolutions:
#note: this isn\t actually the resolution but
# the downsampling factor, i.e. this corresnponds to
# attention on spatial resolution 8,16,32, as the
# spatial reolution of the latents is 32 for f8
- 4
- 2
- 1
num_res_blocks: 2
channel_mult:
- 1
- 2
- 4
num_head_channels: 32
use_spatial_transformer: true
transformer_depth: 1
context_dim: 512
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 4
n_embed: 16384
ckpt_path: configs/first_stage_models/vq-f8/model.yaml
ddconfig:
double_z: false
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 2
- 4
num_res_blocks: 2
attn_resolutions:
- 32
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.ClassEmbedder
params:
embed_dim: 512
key: class_label
data:
target: main.DataModuleFromConfig
params:
batch_size: 64
num_workers: 12
wrap: false
train:
target: ldm.data.imagenet.ImageNetTrain
params:
config:
size: 256
validation:
target: ldm.data.imagenet.ImageNetValidation
params:
config:
size: 256
lightning:
callbacks:
image_logger:
target: main.ImageLogger
params:
batch_frequency: 5000
max_images: 8
increase_log_steps: False
trainer:
benchmark: True

68
configs/latent-diffusion/cin256-v2.yaml Normal file
View File

@ -0,0 +1,68 @@
model:
base_learning_rate: 0.0001
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0195
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: image
cond_stage_key: class_label
image_size: 64
channels: 3
cond_stage_trainable: true
conditioning_key: crossattn
monitor: val/loss
use_ema: False
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 64
in_channels: 3
out_channels: 3
model_channels: 192
attention_resolutions:
- 8
- 4
- 2
num_res_blocks: 2
channel_mult:
- 1
- 2
- 3
- 5
num_heads: 1
use_spatial_transformer: true
transformer_depth: 1
context_dim: 512
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 3
n_embed: 8192
ddconfig:
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.ClassEmbedder
params:
n_classes: 1001
embed_dim: 512
key: class_label

85
configs/latent-diffusion/ffhq-ldm-vq-4.yaml Normal file
View File

@ -0,0 +1,85 @@
model:
base_learning_rate: 2.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0195
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: image
image_size: 64
channels: 3
monitor: val/loss_simple_ema
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 64
in_channels: 3
out_channels: 3
model_channels: 224
attention_resolutions:
# note: this isn\t actually the resolution but
# the downsampling factor, i.e. this corresnponds to
# attention on spatial resolution 8,16,32, as the
# spatial reolution of the latents is 64 for f4
- 8
- 4
- 2
num_res_blocks: 2
channel_mult:
- 1
- 2
- 3
- 4
num_head_channels: 32
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 3
n_embed: 8192
ckpt_path: configs/first_stage_models/vq-f4/model.yaml
ddconfig:
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config: __is_unconditional__
data:
target: main.DataModuleFromConfig
params:
batch_size: 42
num_workers: 5
wrap: false
train:
target: taming.data.faceshq.FFHQTrain
params:
size: 256
validation:
target: taming.data.faceshq.FFHQValidation
params:
size: 256
lightning:
callbacks:
image_logger:
target: main.ImageLogger
params:
batch_frequency: 5000
max_images: 8
increase_log_steps: False
trainer:
benchmark: True

85
configs/latent-diffusion/lsun_bedrooms-ldm-vq-4.yaml Normal file
View File

@ -0,0 +1,85 @@
model:
base_learning_rate: 2.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0195
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: image
image_size: 64
channels: 3
monitor: val/loss_simple_ema
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 64
in_channels: 3
out_channels: 3
model_channels: 224
attention_resolutions:
# note: this isn\t actually the resolution but
# the downsampling factor, i.e. this corresnponds to
# attention on spatial resolution 8,16,32, as the
# spatial reolution of the latents is 64 for f4
- 8
- 4
- 2
num_res_blocks: 2
channel_mult:
- 1
- 2
- 3
- 4
num_head_channels: 32
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
ckpt_path: configs/first_stage_models/vq-f4/model.yaml
embed_dim: 3
n_embed: 8192
ddconfig:
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config: __is_unconditional__
data:
target: main.DataModuleFromConfig
params:
batch_size: 48
num_workers: 5
wrap: false
train:
target: ldm.data.lsun.LSUNBedroomsTrain
params:
size: 256
validation:
target: ldm.data.lsun.LSUNBedroomsValidation
params:
size: 256
lightning:
callbacks:
image_logger:
target: main.ImageLogger
params:
batch_frequency: 5000
max_images: 8
increase_log_steps: False
trainer:
benchmark: True

91
configs/latent-diffusion/lsun_churches-ldm-kl-8.yaml Normal file
View File

@ -0,0 +1,91 @@
model:
base_learning_rate: 5.0e-5 # set to target_lr by starting main.py with '--scale_lr False'
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0155
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
loss_type: l1
first_stage_key: "image"
cond_stage_key: "image"
image_size: 32
channels: 4
cond_stage_trainable: False
concat_mode: False
scale_by_std: True
monitor: 'val/loss_simple_ema'
scheduler_config: # 10000 warmup steps
target: ldm.lr_scheduler.LambdaLinearScheduler
params:
warm_up_steps: [10000]
cycle_lengths: [10000000000000]
f_start: [1.e-6]
f_max: [1.]
f_min: [ 1.]
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 32
in_channels: 4
out_channels: 4
model_channels: 192
attention_resolutions: [ 1, 2, 4, 8 ] # 32, 16, 8, 4
num_res_blocks: 2
channel_mult: [ 1,2,2,4,4 ] # 32, 16, 8, 4, 2
num_heads: 8
use_scale_shift_norm: True
resblock_updown: True
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: "val/rec_loss"
ckpt_path: "models/first_stage_models/kl-f8/model.ckpt"
ddconfig:
double_z: True
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult: [ 1,2,4,4 ] # num_down = len(ch_mult)-1
num_res_blocks: 2
attn_resolutions: [ ]
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config: "__is_unconditional__"
data:
target: main.DataModuleFromConfig
params:
batch_size: 96
num_workers: 5
wrap: False
train:
target: ldm.data.lsun.LSUNChurchesTrain
params:
size: 256
validation:
target: ldm.data.lsun.LSUNChurchesValidation
params:
size: 256
lightning:
callbacks:
image_logger:
target: main.ImageLogger
params:
batch_frequency: 5000
max_images: 8
increase_log_steps: False
trainer:
benchmark: True

71
configs/latent-diffusion/txt2img-1p4B-eval.yaml Normal file
View File

@ -0,0 +1,71 @@
model:
base_learning_rate: 5.0e-05
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.00085
linear_end: 0.012
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: image
cond_stage_key: caption
image_size: 32
channels: 4
cond_stage_trainable: true
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 32
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions:
- 4
- 2
- 1
num_res_blocks: 2
channel_mult:
- 1
- 2
- 4
- 4
num_heads: 8
use_spatial_transformer: true
transformer_depth: 1
context_dim: 1280
use_checkpoint: true
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.BERTEmbedder
params:
n_embed: 1280
n_layer: 32

68
configs/retrieval-augmented-diffusion/768x768.yaml Normal file
View File

@ -0,0 +1,68 @@
model:
base_learning_rate: 0.0001
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.015
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: jpg
cond_stage_key: nix
image_size: 48
channels: 16
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_by_std: false
scale_factor: 0.22765929
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 48
in_channels: 16
out_channels: 16
model_channels: 448
attention_resolutions:
- 4
- 2
- 1
num_res_blocks: 2
channel_mult:
- 1
- 2
- 3
- 4
use_scale_shift_norm: false
resblock_updown: false
num_head_channels: 32
use_spatial_transformer: true
transformer_depth: 1
context_dim: 768
use_checkpoint: true
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
monitor: val/rec_loss
embed_dim: 16
ddconfig:
double_z: true
z_channels: 16
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 1
- 2
- 2
- 4
num_res_blocks: 2
attn_resolutions:
- 16
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: torch.nn.Identity

70
configs/stable-diffusion/v1-inference.yaml Normal file
View File

@ -0,0 +1,70 @@
model:
base_learning_rate: 1.0e-04
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
image_size: 64
channels: 4
cond_stage_trainable: false # Note: different from the one we trained before
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
scheduler_config: # 10000 warmup steps
target: ldm.lr_scheduler.LambdaLinearScheduler
params:
warm_up_steps: [ 10000 ]
cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
f_start: [ 1.e-6 ]
f_max: [ 1. ]
f_min: [ 1. ]
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

19
webui.yaml → configs/webui/webui.yaml
View File

@ -4,16 +4,15 @@ txt2img:
prompt:
ddim_steps: 50
# Adding an int to toggles enables the corresponding feature.
# 0: Create prompt matrix (separate multiple prompts using |, and get all combinations of them)
# 1: Normalize Prompt Weights (ensure sum of weights add up to 1.0)
# 2: Save individual images
# 3: Save grid
# 4: Sort samples by prompt
# 5: Write sample info files
# 6: write sample info to log file
# 7: jpg samples
# 8: Fix faces using GFPGAN
# 9: Upscale images using RealESRGAN
# 0: Create prompt matrix (separate multiple prompts using |, and get all combinations of them)
# 1: Normalize Prompt Weights (ensure sum of weights add up to 1.0)
# 2: Save individual images
# 3: Save grid
# 4: Sort samples by prompt
# 5: Write sample info files
# 6: jpg samples
# 7: Fix faces using GFPGAN
# 8: Upscale images using Real-ESRGAN
toggles: [1, 2, 3, 4, 5]
sampler_name: k_lms
ddim_eta: 0.0 # legacy name, applies to all algorithms.

BIN
data/DejaVuSans.ttf Normal file
View File

Binary file not shown.

25
docker-compose.yml Normal file
View File

@ -0,0 +1,25 @@
version: '3.3'
services:
stable-diffusion:
container_name: sd
build:
context: .
dockerfile: Dockerfile
env_file: .env_docker
volumes:
- .:/sd
- ./outputs:/sd/outputs
- conda_env:/opt/conda
- root_profile:/root
ports:
- '7860:7860'
deploy:
resources:
reservations:
devices:
- capabilities: [gpu]
volumes:
conda_env:
root_profile:

22
docker-reset.sh Normal file
View File

@ -0,0 +1,22 @@
#!/bin/bash
# It basically resets you to the beginning except for your output directory.
# How to:
# cd stable-diffusion
# ./docker-reset.sh
# Then:
# docker-compose up
echo $(pwd)
read -p "Is the directory above correct to run reset on? (y/n) " -n 1 DIRCONFIRM
if [[ $DIRCONFIRM =~ ^[Yy]$ ]]; then
docker compose down
docker image rm stable-diffusion_stable-diffusion:latest
docker volume rm stable-diffusion_conda_env
docker volume rm stable-diffusion_root_profile
echo "Remove ./src"
sudo rm -rf src
sudo rm -rf latent_diffusion.egg-info
sudo rm .env_updated
else
echo "Exited without resetting"
fi

104
entrypoint.sh Normal file
View File

@ -0,0 +1,104 @@
#!/bin/bash
#
# Starts the gui inside the docker container using the conda env
#
# Array of model files to pre-download
# local filename
# local path in container (no trailing slash)
# download URL
# sha256sum
MODEL_FILES=(
'model.ckpt /sd/models/ldm/stable-diffusion-v1 https://www.googleapis.com/storage/v1/b/aai-blog-files/o/sd-v1-4.ckpt?alt=media fe4efff1e174c627256e44ec2991ba279b3816e364b49f9be2abc0b3ff3f8556'
'GFPGANv1.3.pth /sd/src/gfpgan/experiments/pretrained_models https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth c953a88f2727c85c3d9ae72e2bd4846bbaf59fe6972ad94130e23e7017524a70'
'RealESRGAN_x4plus.pth /sd/src/realesrgan/experiments/pretrained_models https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth 4fa0d38905f75ac06eb49a7951b426670021be3018265fd191d2125df9d682f1'
'RealESRGAN_x4plus_anime_6B.pth /sd/src/realesrgan/experiments/pretrained_models https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth f872d837d3c90ed2e05227bed711af5671a6fd1c9f7d7e91c911a61f155e99da'
)
# Conda environment installs/updates
# @see https://github.com/ContinuumIO/docker-images/issues/89#issuecomment-467287039
ENV_NAME="ldm"
ENV_FILE="/sd/environment.yaml"
ENV_UPDATED=0
ENV_MODIFIED=$(date -r $ENV_FILE "+%s")
ENV_MODIFED_FILE="/sd/.env_updated"
if [[ -f $ENV_MODIFED_FILE ]]; then ENV_MODIFIED_CACHED=$(<${ENV_MODIFED_FILE}); else ENV_MODIFIED_CACHED=0; fi
# Create/update conda env if needed
if ! conda env list | grep ".*${ENV_NAME}.*" >/dev/null 2>&1; then
echo "Could not find conda env: ${ENV_NAME} ... creating ..."
conda env create -f $ENV_FILE
echo "source activate ${ENV_NAME}" > /root/.bashrc
ENV_UPDATED=1
elif [[ ! -z $CONDA_FORCE_UPDATE && $CONDA_FORCE_UPDATE == "true" ]] || (( $ENV_MODIFIED > $ENV_MODIFIED_CACHED )); then
echo "Updating conda env: ${ENV_NAME} ..."
conda env update --file $ENV_FILE --prune
ENV_UPDATED=1
fi
# Clear artifacts from conda after create/update
# @see https://docs.conda.io/projects/conda/en/latest/commands/clean.html
if (( $ENV_UPDATED > 0 )); then
conda clean --all
echo -n $ENV_MODIFIED > $ENV_MODIFED_FILE
fi
# activate conda env
. /opt/conda/etc/profile.d/conda.sh
conda activate $ENV_NAME
conda info | grep active
# Function to checks for valid hash for model files and download/replaces if invalid or does not exist
validateDownloadModel() {
local file=$1
local path=$2
local url=$3
local hash=$4
echo "checking ${file}..."
sha256sum --check --status <<< "${hash} ${path}/${file}"
if [[ $? == "1" ]]; then
echo "Downloading: ${url} please wait..."
mkdir -p ${path}
wget --output-document=${path}/${file} --no-verbose --show-progress --progress=dot:giga ${url}
echo "saved ${file}"
else
echo -e "${file} is valid!\n"
fi
}
# Validate model files
if [[ -z $VALIDATE_MODELS || $VALIDATE_MODELS == "true" ]]; then
echo "Validating model files..."
for models in "${MODEL_FILES[@]}"; do
model=($models)
validateDownloadModel ${model[0]} ${model[1]} ${model[2]} ${model[3]}
done
fi
# Launch web gui
cd /sd
if [[ -z $WEBUI_ARGS ]]; then
launch_message="entrypoint.sh: Launching..."
else
launch_message="entrypoint.sh: Launching with arguments ${WEBUI_ARGS}"
fi
if [[ -z $WEBUI_RELAUNCH || $WEBUI_RELAUNCH == "true" ]]; then
n=0
while true; do
echo $launch_message
if (( $n > 0 )); then
echo "Relaunch count: ${n}"
fi
python -u scripts/webui.py $WEBUI_ARGS
echo "entrypoint.sh: Process is ending. Relaunching in 0.5s..."
((n++))
sleep 0.5
done
else
echo $launch_message
python -u scripts/webui.py $WEBUI_ARGS
fi

41
environment.yaml Normal file
View File

@ -0,0 +1,41 @@
name: ldm
channels:
- pytorch
- defaults
dependencies:
- git
- python=3.8.5
- pip=20.3
- cudatoolkit=11.3
- pytorch=1.11.0
- torchvision=0.12.0
- numpy=1.19.2
- pip:
- albumentations==0.4.3
- opencv-python==4.1.2.30
- opencv-python-headless==4.1.2.30
- pudb==2019.2
- imageio==2.9.0
- imageio-ffmpeg==0.4.2
- pytorch-lightning==1.4.2
- omegaconf==2.1.1
- test-tube>=0.7.5
- streamlit>=0.73.1
- einops==0.3.0
- torch-fidelity==0.3.0
- transformers==4.19.2
- torchmetrics==0.6.0
- kornia==0.6
- gradio==3.1.6
- accelerate==0.12.0
- pynvml==11.4.1
- basicsr>=1.3.4.0
- facexlib>=0.2.3
- python-slugify>=6.1.2
- streamlit>=1.12.2
- taming-transformers>=0.0.1
- openai-clip>=1.0.1
- gfpgan>=1.3.5
- realesrgan>=0.2.5.0
- -e git+https://github.com/crowsonkb/k-diffusion#egg=k_diffusion
- -e .

0
ldm/data/__init__.py Normal file
View File

23
ldm/data/base.py Normal file
View File

@ -0,0 +1,23 @@
from abc import abstractmethod
from torch.utils.data import Dataset, ConcatDataset, ChainDataset, IterableDataset
class Txt2ImgIterableBaseDataset(IterableDataset):
'''
Define an interface to make the IterableDatasets for text2img data chainable
'''
def __init__(self, num_records=0, valid_ids=None, size=256):
super().__init__()
self.num_records = num_records
self.valid_ids = valid_ids
self.sample_ids = valid_ids
self.size = size
print(f'{self.__class__.__name__} dataset contains {self.__len__()} examples.')
def __len__(self):
return self.num_records
@abstractmethod
def __iter__(self):
pass

394
ldm/data/imagenet.py Normal file
View File

@ -0,0 +1,394 @@
import os, yaml, pickle, shutil, tarfile, glob
import cv2
import albumentations
import PIL
import numpy as np
import torchvision.transforms.functional as TF
from omegaconf import OmegaConf
from functools import partial
from PIL import Image
from tqdm import tqdm
from torch.utils.data import Dataset, Subset
import taming.data.utils as tdu
from taming.data.imagenet import str_to_indices, give_synsets_from_indices, download, retrieve
from taming.data.imagenet import ImagePaths
from ldm.modules.image_degradation import degradation_fn_bsr, degradation_fn_bsr_light
def synset2idx(path_to_yaml="data/index_synset.yaml"):
with open(path_to_yaml) as f:
di2s = yaml.load(f)
return dict((v,k) for k,v in di2s.items())
class ImageNetBase(Dataset):
def __init__(self, config=None):
self.config = config or OmegaConf.create()
if not type(self.config)==dict:
self.config = OmegaConf.to_container(self.config)
self.keep_orig_class_label = self.config.get("keep_orig_class_label", False)
self.process_images = True # if False we skip loading & processing images and self.data contains filepaths
self._prepare()
self._prepare_synset_to_human()
self._prepare_idx_to_synset()
self._prepare_human_to_integer_label()
self._load()
def __len__(self):
return len(self.data)
def __getitem__(self, i):
return self.data[i]
def _prepare(self):
raise NotImplementedError()
def _filter_relpaths(self, relpaths):
ignore = set([
"n06596364_9591.JPEG",
])
relpaths = [rpath for rpath in relpaths if not rpath.split("/")[-1] in ignore]
if "sub_indices" in self.config:
indices = str_to_indices(self.config["sub_indices"])
synsets = give_synsets_from_indices(indices, path_to_yaml=self.idx2syn) # returns a list of strings
self.synset2idx = synset2idx(path_to_yaml=self.idx2syn)
files = []
for rpath in relpaths:
syn = rpath.split("/")[0]
if syn in synsets:
files.append(rpath)
return files
else:
return relpaths
def _prepare_synset_to_human(self):
SIZE = 2655750
URL = "https://heibox.uni-heidelberg.de/f/9f28e956cd304264bb82/?dl=1"
self.human_dict = os.path.join(self.root, "synset_human.txt")
if (not os.path.exists(self.human_dict) or
not os.path.getsize(self.human_dict)==SIZE):
download(URL, self.human_dict)
def _prepare_idx_to_synset(self):
URL = "https://heibox.uni-heidelberg.de/f/d835d5b6ceda4d3aa910/?dl=1"
self.idx2syn = os.path.join(self.root, "index_synset.yaml")
if (not os.path.exists(self.idx2syn)):
download(URL, self.idx2syn)
def _prepare_human_to_integer_label(self):
URL = "https://heibox.uni-heidelberg.de/f/2362b797d5be43b883f6/?dl=1"
self.human2integer = os.path.join(self.root, "imagenet1000_clsidx_to_labels.txt")
if (not os.path.exists(self.human2integer)):
download(URL, self.human2integer)
with open(self.human2integer, "r") as f:
lines = f.read().splitlines()
assert len(lines) == 1000
self.human2integer_dict = dict()
for line in lines:
value, key = line.split(":")
self.human2integer_dict[key] = int(value)
def _load(self):
with open(self.txt_filelist, "r") as f:
self.relpaths = f.read().splitlines()
l1 = len(self.relpaths)
self.relpaths = self._filter_relpaths(self.relpaths)
print("Removed {} files from filelist during filtering.".format(l1 - len(self.relpaths)))
self.synsets = [p.split("/")[0] for p in self.relpaths]
self.abspaths = [os.path.join(self.datadir, p) for p in self.relpaths]
unique_synsets = np.unique(self.synsets)
class_dict = dict((synset, i) for i, synset in enumerate(unique_synsets))
if not self.keep_orig_class_label:
self.class_labels = [class_dict[s] for s in self.synsets]
else:
self.class_labels = [self.synset2idx[s] for s in self.synsets]
with open(self.human_dict, "r") as f:
human_dict = f.read().splitlines()
human_dict = dict(line.split(maxsplit=1) for line in human_dict)
self.human_labels = [human_dict[s] for s in self.synsets]
labels = {
"relpath": np.array(self.relpaths),
"synsets": np.array(self.synsets),
"class_label": np.array(self.class_labels),
"human_label": np.array(self.human_labels),
}
if self.process_images:
self.size = retrieve(self.config, "size", default=256)
self.data = ImagePaths(self.abspaths,
labels=labels,
size=self.size,
random_crop=self.random_crop,
)
else:
self.data = self.abspaths
class ImageNetTrain(ImageNetBase):
NAME = "ILSVRC2012_train"
URL = "http://www.image-net.org/challenges/LSVRC/2012/"
AT_HASH = "a306397ccf9c2ead27155983c254227c0fd938e2"
FILES = [
"ILSVRC2012_img_train.tar",
]
SIZES = [
147897477120,
]
def __init__(self, process_images=True, data_root=None, **kwargs):
self.process_images = process_images
self.data_root = data_root
super().__init__(**kwargs)
def _prepare(self):
if self.data_root:
self.root = os.path.join(self.data_root, self.NAME)
else:
cachedir = os.environ.get("XDG_CACHE_HOME", os.path.expanduser("~/.cache"))
self.root = os.path.join(cachedir, "autoencoders/data", self.NAME)
self.datadir = os.path.join(self.root, "data")
self.txt_filelist = os.path.join(self.root, "filelist.txt")
self.expected_length = 1281167
self.random_crop = retrieve(self.config, "ImageNetTrain/random_crop",
default=True)
if not tdu.is_prepared(self.root):
# prep
print("Preparing dataset {} in {}".format(self.NAME, self.root))
datadir = self.datadir
if not os.path.exists(datadir):
path = os.path.join(self.root, self.FILES[0])
if not os.path.exists(path) or not os.path.getsize(path)==self.SIZES[0]:
import academictorrents as at
atpath = at.get(self.AT_HASH, datastore=self.root)
assert atpath == path
print("Extracting {} to {}".format(path, datadir))
os.makedirs(datadir, exist_ok=True)
with tarfile.open(path, "r:") as tar:
tar.extractall(path=datadir)
print("Extracting sub-tars.")
subpaths = sorted(glob.glob(os.path.join(datadir, "*.tar")))
for subpath in tqdm(subpaths):
subdir = subpath[:-len(".tar")]
os.makedirs(subdir, exist_ok=True)
with tarfile.open(subpath, "r:") as tar:
tar.extractall(path=subdir)
filelist = glob.glob(os.path.join(datadir, "**", "*.JPEG"))
filelist = [os.path.relpath(p, start=datadir) for p in filelist]
filelist = sorted(filelist)
filelist = "\n".join(filelist)+"\n"
with open(self.txt_filelist, "w") as f:
f.write(filelist)
tdu.mark_prepared(self.root)
class ImageNetValidation(ImageNetBase):
NAME = "ILSVRC2012_validation"
URL = "http://www.image-net.org/challenges/LSVRC/2012/"
AT_HASH = "5d6d0df7ed81efd49ca99ea4737e0ae5e3a5f2e5"
VS_URL = "https://heibox.uni-heidelberg.de/f/3e0f6e9c624e45f2bd73/?dl=1"
FILES = [
"ILSVRC2012_img_val.tar",
"validation_synset.txt",
]
SIZES = [
6744924160,
1950000,
]
def __init__(self, process_images=True, data_root=None, **kwargs):
self.data_root = data_root
self.process_images = process_images
super().__init__(**kwargs)
def _prepare(self):
if self.data_root:
self.root = os.path.join(self.data_root, self.NAME)
else:
cachedir = os.environ.get("XDG_CACHE_HOME", os.path.expanduser("~/.cache"))
self.root = os.path.join(cachedir, "autoencoders/data", self.NAME)
self.datadir = os.path.join(self.root, "data")
self.txt_filelist = os.path.join(self.root, "filelist.txt")
self.expected_length = 50000
self.random_crop = retrieve(self.config, "ImageNetValidation/random_crop",
default=False)
if not tdu.is_prepared(self.root):
# prep
print("Preparing dataset {} in {}".format(self.NAME, self.root))
datadir = self.datadir
if not os.path.exists(datadir):
path = os.path.join(self.root, self.FILES[0])
if not os.path.exists(path) or not os.path.getsize(path)==self.SIZES[0]:
import academictorrents as at
atpath = at.get(self.AT_HASH, datastore=self.root)
assert atpath == path
print("Extracting {} to {}".format(path, datadir))
os.makedirs(datadir, exist_ok=True)
with tarfile.open(path, "r:") as tar:
tar.extractall(path=datadir)
vspath = os.path.join(self.root, self.FILES[1])
if not os.path.exists(vspath) or not os.path.getsize(vspath)==self.SIZES[1]:
download(self.VS_URL, vspath)
with open(vspath, "r") as f:
synset_dict = f.read().splitlines()
synset_dict = dict(line.split() for line in synset_dict)
print("Reorganizing into synset folders")
synsets = np.unique(list(synset_dict.values()))
for s in synsets:
os.makedirs(os.path.join(datadir, s), exist_ok=True)
for k, v in synset_dict.items():
src = os.path.join(datadir, k)
dst = os.path.join(datadir, v)
shutil.move(src, dst)
filelist = glob.glob(os.path.join(datadir, "**", "*.JPEG"))
filelist = [os.path.relpath(p, start=datadir) for p in filelist]
filelist = sorted(filelist)
filelist = "\n".join(filelist)+"\n"
with open(self.txt_filelist, "w") as f:
f.write(filelist)
tdu.mark_prepared(self.root)
class ImageNetSR(Dataset):
def __init__(self, size=None,
degradation=None, downscale_f=4, min_crop_f=0.5, max_crop_f=1.,
random_crop=True):
"""
Imagenet Superresolution Dataloader
Performs following ops in order:
1. crops a crop of size s from image either as random or center crop
2. resizes crop to size with cv2.area_interpolation
3. degrades resized crop with degradation_fn
:param size: resizing to size after cropping
:param degradation: degradation_fn, e.g. cv_bicubic or bsrgan_light
:param downscale_f: Low Resolution Downsample factor
:param min_crop_f: determines crop size s,
where s = c * min_img_side_len with c sampled from interval (min_crop_f, max_crop_f)
:param max_crop_f: ""
:param data_root:
:param random_crop:
"""
self.base = self.get_base()
assert size
assert (size / downscale_f).is_integer()
self.size = size
self.LR_size = int(size / downscale_f)
self.min_crop_f = min_crop_f
self.max_crop_f = max_crop_f
assert(max_crop_f <= 1.)
self.center_crop = not random_crop
self.image_rescaler = albumentations.SmallestMaxSize(max_size=size, interpolation=cv2.INTER_AREA)
self.pil_interpolation = False # gets reset later if incase interp_op is from pillow
if degradation == "bsrgan":
self.degradation_process = partial(degradation_fn_bsr, sf=downscale_f)
elif degradation == "bsrgan_light":
self.degradation_process = partial(degradation_fn_bsr_light, sf=downscale_f)
else:
interpolation_fn = {
"cv_nearest": cv2.INTER_NEAREST,
"cv_bilinear": cv2.INTER_LINEAR,
"cv_bicubic": cv2.INTER_CUBIC,
"cv_area": cv2.INTER_AREA,
"cv_lanczos": cv2.INTER_LANCZOS4,
"pil_nearest": PIL.Image.NEAREST,
"pil_bilinear": PIL.Image.BILINEAR,
"pil_bicubic": PIL.Image.BICUBIC,
"pil_box": PIL.Image.BOX,
"pil_hamming": PIL.Image.HAMMING,
"pil_lanczos": PIL.Image.LANCZOS,
}[degradation]
self.pil_interpolation = degradation.startswith("pil_")
if self.pil_interpolation:
self.degradation_process = partial(TF.resize, size=self.LR_size, interpolation=interpolation_fn)
else:
self.degradation_process = albumentations.SmallestMaxSize(max_size=self.LR_size,
interpolation=interpolation_fn)
def __len__(self):
return len(self.base)
def __getitem__(self, i):
example = self.base[i]
image = Image.open(example["file_path_"])
if not image.mode == "RGB":
image = image.convert("RGB")
image = np.array(image).astype(np.uint8)
min_side_len = min(image.shape[:2])
crop_side_len = min_side_len * np.random.uniform(self.min_crop_f, self.max_crop_f, size=None)
crop_side_len = int(crop_side_len)
if self.center_crop:
self.cropper = albumentations.CenterCrop(height=crop_side_len, width=crop_side_len)
else:
self.cropper = albumentations.RandomCrop(height=crop_side_len, width=crop_side_len)
image = self.cropper(image=image)["image"]
image = self.image_rescaler(image=image)["image"]
if self.pil_interpolation:
image_pil = PIL.Image.fromarray(image)
LR_image = self.degradation_process(image_pil)
LR_image = np.array(LR_image).astype(np.uint8)
else:
LR_image = self.degradation_process(image=image)["image"]
example["image"] = (image/127.5 - 1.0).astype(np.float32)
example["LR_image"] = (LR_image/127.5 - 1.0).astype(np.float32)
return example
class ImageNetSRTrain(ImageNetSR):
def __init__(self, **kwargs):
super().__init__(**kwargs)
def get_base(self):
with open("data/imagenet_train_hr_indices.p", "rb") as f:
indices = pickle.load(f)
dset = ImageNetTrain(process_images=False,)
return Subset(dset, indices)
class ImageNetSRValidation(ImageNetSR):
def __init__(self, **kwargs):
super().__init__(**kwargs)
def get_base(self):
with open("data/imagenet_val_hr_indices.p", "rb") as f:
indices = pickle.load(f)
dset = ImageNetValidation(process_images=False,)
return Subset(dset, indices)

92
ldm/data/lsun.py Normal file
View File

@ -0,0 +1,92 @@
import os
import numpy as np
import PIL
from PIL import Image
from torch.utils.data import Dataset
from torchvision import transforms
class LSUNBase(Dataset):
def __init__(self,
txt_file,
data_root,
size=None,
interpolation="bicubic",
flip_p=0.5
):
self.data_paths = txt_file
self.data_root = data_root
with open(self.data_paths, "r") as f:
self.image_paths = f.read().splitlines()
self._length = len(self.image_paths)
self.labels = {
"relative_file_path_": [l for l in self.image_paths],
"file_path_": [os.path.join(self.data_root, l)
for l in self.image_paths],
}
self.size = size
self.interpolation = {"linear": PIL.Image.LINEAR,
"bilinear": PIL.Image.BILINEAR,
"bicubic": PIL.Image.BICUBIC,
"lanczos": PIL.Image.LANCZOS,
}[interpolation]
self.flip = transforms.RandomHorizontalFlip(p=flip_p)
def __len__(self):
return self._length
def __getitem__(self, i):
example = dict((k, self.labels[k][i]) for k in self.labels)
image = Image.open(example["file_path_"])
if not image.mode == "RGB":
image = image.convert("RGB")
# default to score-sde preprocessing
img = np.array(image).astype(np.uint8)
crop = min(img.shape[0], img.shape[1])
h, w, = img.shape[0], img.shape[1]
img = img[(h - crop) // 2:(h + crop) // 2,
(w - crop) // 2:(w + crop) // 2]
image = Image.fromarray(img)
if self.size is not None:
image = image.resize((self.size, self.size), resample=self.interpolation)
image = self.flip(image)
image = np.array(image).astype(np.uint8)
example["image"] = (image / 127.5 - 1.0).astype(np.float32)
return example
class LSUNChurchesTrain(LSUNBase):
def __init__(self, **kwargs):
super().__init__(txt_file="data/lsun/church_outdoor_train.txt", data_root="data/lsun/churches", **kwargs)
class LSUNChurchesValidation(LSUNBase):
def __init__(self, flip_p=0., **kwargs):
super().__init__(txt_file="data/lsun/church_outdoor_val.txt", data_root="data/lsun/churches",
flip_p=flip_p, **kwargs)
class LSUNBedroomsTrain(LSUNBase):
def __init__(self, **kwargs):
super().__init__(txt_file="data/lsun/bedrooms_train.txt", data_root="data/lsun/bedrooms", **kwargs)
class LSUNBedroomsValidation(LSUNBase):
def __init__(self, flip_p=0.0, **kwargs):
super().__init__(txt_file="data/lsun/bedrooms_val.txt", data_root="data/lsun/bedrooms",
flip_p=flip_p, **kwargs)
class LSUNCatsTrain(LSUNBase):
def __init__(self, **kwargs):
super().__init__(txt_file="data/lsun/cat_train.txt", data_root="data/lsun/cats", **kwargs)
class LSUNCatsValidation(LSUNBase):
def __init__(self, flip_p=0., **kwargs):
super().__init__(txt_file="data/lsun/cat_val.txt", data_root="data/lsun/cats",
flip_p=flip_p, **kwargs)

98
ldm/lr_scheduler.py Normal file
View File

@ -0,0 +1,98 @@
import numpy as np
class LambdaWarmUpCosineScheduler:
"""
note: use with a base_lr of 1.0
"""
def __init__(self, warm_up_steps, lr_min, lr_max, lr_start, max_decay_steps, verbosity_interval=0):
self.lr_warm_up_steps = warm_up_steps
self.lr_start = lr_start
self.lr_min = lr_min
self.lr_max = lr_max
self.lr_max_decay_steps = max_decay_steps
self.last_lr = 0.
self.verbosity_interval = verbosity_interval
def schedule(self, n, **kwargs):
if self.verbosity_interval > 0:
if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_lr}")
if n < self.lr_warm_up_steps:
lr = (self.lr_max - self.lr_start) / self.lr_warm_up_steps * n + self.lr_start
self.last_lr = lr
return lr
else:
t = (n - self.lr_warm_up_steps) / (self.lr_max_decay_steps - self.lr_warm_up_steps)
t = min(t, 1.0)
lr = self.lr_min + 0.5 * (self.lr_max - self.lr_min) * (
1 + np.cos(t * np.pi))
self.last_lr = lr
return lr
def __call__(self, n, **kwargs):
return self.schedule(n,**kwargs)
class LambdaWarmUpCosineScheduler2:
"""
supports repeated iterations, configurable via lists
note: use with a base_lr of 1.0.
"""
def __init__(self, warm_up_steps, f_min, f_max, f_start, cycle_lengths, verbosity_interval=0):
assert len(warm_up_steps) == len(f_min) == len(f_max) == len(f_start) == len(cycle_lengths)
self.lr_warm_up_steps = warm_up_steps
self.f_start = f_start
self.f_min = f_min
self.f_max = f_max
self.cycle_lengths = cycle_lengths
self.cum_cycles = np.cumsum([0] + list(self.cycle_lengths))
self.last_f = 0.
self.verbosity_interval = verbosity_interval
def find_in_interval(self, n):
interval = 0
for cl in self.cum_cycles[1:]:
if n <= cl:
return interval
interval += 1
def schedule(self, n, **kwargs):
cycle = self.find_in_interval(n)
n = n - self.cum_cycles[cycle]
if self.verbosity_interval > 0:
if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
f"current cycle {cycle}")
if n < self.lr_warm_up_steps[cycle]:
f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
self.last_f = f
return f
else:
t = (n - self.lr_warm_up_steps[cycle]) / (self.cycle_lengths[cycle] - self.lr_warm_up_steps[cycle])
t = min(t, 1.0)
f = self.f_min[cycle] + 0.5 * (self.f_max[cycle] - self.f_min[cycle]) * (
1 + np.cos(t * np.pi))
self.last_f = f
return f
def __call__(self, n, **kwargs):
return self.schedule(n, **kwargs)
class LambdaLinearScheduler(LambdaWarmUpCosineScheduler2):
def schedule(self, n, **kwargs):
cycle = self.find_in_interval(n)
n = n - self.cum_cycles[cycle]
if self.verbosity_interval > 0:
if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
f"current cycle {cycle}")
if n < self.lr_warm_up_steps[cycle]:
f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
self.last_f = f
return f
else:
f = self.f_min[cycle] + (self.f_max[cycle] - self.f_min[cycle]) * (self.cycle_lengths[cycle] - n) / (self.cycle_lengths[cycle])
self.last_f = f
return f

443
ldm/models/autoencoder.py Normal file
View File

@ -0,0 +1,443 @@
import torch
import pytorch_lightning as pl
import torch.nn.functional as F
from contextlib import contextmanager
from taming.modules.vqvae.quantize import VectorQuantizer2 as VectorQuantizer
from ldm.modules.diffusionmodules.model import Encoder, Decoder
from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
from ldm.util import instantiate_from_config
class VQModel(pl.LightningModule):
def __init__(self,
ddconfig,
lossconfig,
n_embed,
embed_dim,
ckpt_path=None,
ignore_keys=[],
image_key="image",
colorize_nlabels=None,
monitor=None,
batch_resize_range=None,
scheduler_config=None,
lr_g_factor=1.0,
remap=None,
sane_index_shape=False, # tell vector quantizer to return indices as bhw
use_ema=False
):
super().__init__()
self.embed_dim = embed_dim
self.n_embed = n_embed
self.image_key = image_key
self.encoder = Encoder(**ddconfig)
self.decoder = Decoder(**ddconfig)
self.loss = instantiate_from_config(lossconfig)
self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25,
remap=remap,
sane_index_shape=sane_index_shape)
self.quant_conv = torch.nn.Conv2d(ddconfig["z_channels"], embed_dim, 1)
self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
if colorize_nlabels is not None:
assert type(colorize_nlabels)==int
self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
if monitor is not None:
self.monitor = monitor
self.batch_resize_range = batch_resize_range
if self.batch_resize_range is not None:
print(f"{self.__class__.__name__}: Using per-batch resizing in range {batch_resize_range}.")
self.use_ema = use_ema
if self.use_ema:
self.model_ema = LitEma(self)
print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
if ckpt_path is not None:
self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
self.scheduler_config = scheduler_config
self.lr_g_factor = lr_g_factor
@contextmanager
def ema_scope(self, context=None):
if self.use_ema:
self.model_ema.store(self.parameters())
self.model_ema.copy_to(self)
if context is not None:
print(f"{context}: Switched to EMA weights")
try:
yield None
finally:
if self.use_ema:
self.model_ema.restore(self.parameters())
if context is not None:
print(f"{context}: Restored training weights")
def init_from_ckpt(self, path, ignore_keys=list()):
sd = torch.load(path, map_location="cpu")["state_dict"]
keys = list(sd.keys())
for k in keys:
for ik in ignore_keys:
if k.startswith(ik):
print("Deleting key {} from state_dict.".format(k))
del sd[k]
missing, unexpected = self.load_state_dict(sd, strict=False)
print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
if len(missing) > 0:
print(f"Missing Keys: {missing}")
print(f"Unexpected Keys: {unexpected}")
def on_train_batch_end(self, *args, **kwargs):
if self.use_ema:
self.model_ema(self)
def encode(self, x):
h = self.encoder(x)
h = self.quant_conv(h)
quant, emb_loss, info = self.quantize(h)
return quant, emb_loss, info
def encode_to_prequant(self, x):
h = self.encoder(x)
h = self.quant_conv(h)
return h
def decode(self, quant):
quant = self.post_quant_conv(quant)
dec = self.decoder(quant)
return dec
def decode_code(self, code_b):
quant_b = self.quantize.embed_code(code_b)
dec = self.decode(quant_b)
return dec
def forward(self, input, return_pred_indices=False):
quant, diff, (_,_,ind) = self.encode(input)
dec = self.decode(quant)
if return_pred_indices:
return dec, diff, ind
return dec, diff
def get_input(self, batch, k):
x = batch[k]
if len(x.shape) == 3:
x = x[..., None]
x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
if self.batch_resize_range is not None:
lower_size = self.batch_resize_range[0]
upper_size = self.batch_resize_range[1]
if self.global_step <= 4:
# do the first few batches with max size to avoid later oom
new_resize = upper_size
else:
new_resize = np.random.choice(np.arange(lower_size, upper_size+16, 16))
if new_resize != x.shape[2]:
x = F.interpolate(x, size=new_resize, mode="bicubic")
x = x.detach()
return x
def training_step(self, batch, batch_idx, optimizer_idx):
# https://github.com/pytorch/pytorch/issues/37142
# try not to fool the heuristics
x = self.get_input(batch, self.image_key)
xrec, qloss, ind = self(x, return_pred_indices=True)
if optimizer_idx == 0:
# autoencode
aeloss, log_dict_ae = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
last_layer=self.get_last_layer(), split="train",
predicted_indices=ind)
self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True)
return aeloss
if optimizer_idx == 1:
# discriminator
discloss, log_dict_disc = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
last_layer=self.get_last_layer(), split="train")
self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True)
return discloss
def validation_step(self, batch, batch_idx):
log_dict = self._validation_step(batch, batch_idx)
with self.ema_scope():
log_dict_ema = self._validation_step(batch, batch_idx, suffix="_ema")
return log_dict
def _validation_step(self, batch, batch_idx, suffix=""):
x = self.get_input(batch, self.image_key)
xrec, qloss, ind = self(x, return_pred_indices=True)
aeloss, log_dict_ae = self.loss(qloss, x, xrec, 0,
self.global_step,
last_layer=self.get_last_layer(),
split="val"+suffix,
predicted_indices=ind
)
discloss, log_dict_disc = self.loss(qloss, x, xrec, 1,
self.global_step,
last_layer=self.get_last_layer(),
split="val"+suffix,
predicted_indices=ind
)
rec_loss = log_dict_ae[f"val{suffix}/rec_loss"]
self.log(f"val{suffix}/rec_loss", rec_loss,
prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
self.log(f"val{suffix}/aeloss", aeloss,
prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
if version.parse(pl.__version__) >= version.parse('1.4.0'):
del log_dict_ae[f"val{suffix}/rec_loss"]
self.log_dict(log_dict_ae)
self.log_dict(log_dict_disc)
return self.log_dict
def configure_optimizers(self):
lr_d = self.learning_rate
lr_g = self.lr_g_factor*self.learning_rate
print("lr_d", lr_d)
print("lr_g", lr_g)
opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
list(self.decoder.parameters())+
list(self.quantize.parameters())+
list(self.quant_conv.parameters())+
list(self.post_quant_conv.parameters()),
lr=lr_g, betas=(0.5, 0.9))
opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
lr=lr_d, betas=(0.5, 0.9))
if self.scheduler_config is not None:
scheduler = instantiate_from_config(self.scheduler_config)
print("Setting up LambdaLR scheduler...")
scheduler = [
{
'scheduler': LambdaLR(opt_ae, lr_lambda=scheduler.schedule),
'interval': 'step',
'frequency': 1
},
{
'scheduler': LambdaLR(opt_disc, lr_lambda=scheduler.schedule),
'interval': 'step',
'frequency': 1
},
]
return [opt_ae, opt_disc], scheduler
return [opt_ae, opt_disc], []
def get_last_layer(self):
return self.decoder.conv_out.weight
def log_images(self, batch, only_inputs=False, plot_ema=False, **kwargs):
log = dict()
x = self.get_input(batch, self.image_key)
x = x.to(self.device)
if only_inputs:
log["inputs"] = x
return log
xrec, _ = self(x)
if x.shape[1] > 3:
# colorize with random projection
assert xrec.shape[1] > 3
x = self.to_rgb(x)
xrec = self.to_rgb(xrec)
log["inputs"] = x
log["reconstructions"] = xrec
if plot_ema:
with self.ema_scope():
xrec_ema, _ = self(x)
if x.shape[1] > 3: xrec_ema = self.to_rgb(xrec_ema)
log["reconstructions_ema"] = xrec_ema
return log
def to_rgb(self, x):
assert self.image_key == "segmentation"
if not hasattr(self, "colorize"):
self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
x = F.conv2d(x, weight=self.colorize)
x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
return x
class VQModelInterface(VQModel):
def __init__(self, embed_dim, *args, **kwargs):
super().__init__(embed_dim=embed_dim, *args, **kwargs)
self.embed_dim = embed_dim
def encode(self, x):
h = self.encoder(x)
h = self.quant_conv(h)
return h
def decode(self, h, force_not_quantize=False):
# also go through quantization layer
if not force_not_quantize:
quant, emb_loss, info = self.quantize(h)
else:
quant = h
quant = self.post_quant_conv(quant)
dec = self.decoder(quant)
return dec
class AutoencoderKL(pl.LightningModule):
def __init__(self,
ddconfig,
lossconfig,
embed_dim,
ckpt_path=None,
ignore_keys=[],
image_key="image",
colorize_nlabels=None,
monitor=None,
):
super().__init__()
self.image_key = image_key
self.encoder = Encoder(**ddconfig)
self.decoder = Decoder(**ddconfig)
self.loss = instantiate_from_config(lossconfig)
assert ddconfig["double_z"]
self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
self.embed_dim = embed_dim
if colorize_nlabels is not None:
assert type(colorize_nlabels)==int
self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
if monitor is not None:
self.monitor = monitor
if ckpt_path is not None:
self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
def init_from_ckpt(self, path, ignore_keys=list()):
sd = torch.load(path, map_location="cpu")["state_dict"]
keys = list(sd.keys())
for k in keys:
for ik in ignore_keys:
if k.startswith(ik):
print("Deleting key {} from state_dict.".format(k))
del sd[k]
self.load_state_dict(sd, strict=False)
print(f"Restored from {path}")
def encode(self, x):
h = self.encoder(x)
moments = self.quant_conv(h)
posterior = DiagonalGaussianDistribution(moments)
return posterior
def decode(self, z):
z = self.post_quant_conv(z)
dec = self.decoder(z)
return dec
def forward(self, input, sample_posterior=True):
posterior = self.encode(input)
if sample_posterior:
z = posterior.sample()
else:
z = posterior.mode()
dec = self.decode(z)
return dec, posterior
def get_input(self, batch, k):
x = batch[k]
if len(x.shape) == 3:
x = x[..., None]
x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
return x
def training_step(self, batch, batch_idx, optimizer_idx):
inputs = self.get_input(batch, self.image_key)
reconstructions, posterior = self(inputs)
if optimizer_idx == 0:
# train encoder+decoder+logvar
aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
last_layer=self.get_last_layer(), split="train")
self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False)
return aeloss
if optimizer_idx == 1:
# train the discriminator
discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
last_layer=self.get_last_layer(), split="train")
self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False)
return discloss
def validation_step(self, batch, batch_idx):
inputs = self.get_input(batch, self.image_key)
reconstructions, posterior = self(inputs)
aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step,
last_layer=self.get_last_layer(), split="val")
discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step,
last_layer=self.get_last_layer(), split="val")
self.log("val/rec_loss", log_dict_ae["val/rec_loss"])
self.log_dict(log_dict_ae)
self.log_dict(log_dict_disc)
return self.log_dict
def configure_optimizers(self):
lr = self.learning_rate
opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
list(self.decoder.parameters())+
list(self.quant_conv.parameters())+
list(self.post_quant_conv.parameters()),
lr=lr, betas=(0.5, 0.9))
opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
lr=lr, betas=(0.5, 0.9))
return [opt_ae, opt_disc], []
def get_last_layer(self):
return self.decoder.conv_out.weight
@torch.no_grad()
def log_images(self, batch, only_inputs=False, **kwargs):
log = dict()
x = self.get_input(batch, self.image_key)
x = x.to(self.device)
if not only_inputs:
xrec, posterior = self(x)
if x.shape[1] > 3:
# colorize with random projection
assert xrec.shape[1] > 3
x = self.to_rgb(x)
xrec = self.to_rgb(xrec)
log["samples"] = self.decode(torch.randn_like(posterior.sample()))
log["reconstructions"] = xrec
log["inputs"] = x
return log
def to_rgb(self, x):
assert self.image_key == "segmentation"
if not hasattr(self, "colorize"):
self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
x = F.conv2d(x, weight=self.colorize)
x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
return x
class IdentityFirstStage(torch.nn.Module):
def __init__(self, *args, vq_interface=False, **kwargs):
self.vq_interface = vq_interface # TODO: Should be true by default but check to not break older stuff
super().__init__()
def encode(self, x, *args, **kwargs):
return x
def decode(self, x, *args, **kwargs):
return x
def quantize(self, x, *args, **kwargs):
if self.vq_interface:
return x, None, [None, None, None]
return x
def forward(self, x, *args, **kwargs):
return x

0
ldm/models/diffusion/__init__.py Normal file
View File

267
ldm/models/diffusion/classifier.py Normal file
View File

@ -0,0 +1,267 @@
import os
import torch
import pytorch_lightning as pl
from omegaconf import OmegaConf
from torch.nn import functional as F
from torch.optim import AdamW
from torch.optim.lr_scheduler import LambdaLR
from copy import deepcopy
from einops import rearrange
from glob import glob
from natsort import natsorted
from ldm.modules.diffusionmodules.openaimodel import EncoderUNetModel, UNetModel
from ldm.util import log_txt_as_img, default, ismap, instantiate_from_config
__models__ = {
'class_label': EncoderUNetModel,
'segmentation': UNetModel
}
def disabled_train(self, mode=True):
"""Overwrite model.train with this function to make sure train/eval mode
does not change anymore."""
return self
class NoisyLatentImageClassifier(pl.LightningModule):
def __init__(self,
diffusion_path,
num_classes,
ckpt_path=None,
pool='attention',
label_key=None,
diffusion_ckpt_path=None,
scheduler_config=None,
weight_decay=1.e-2,
log_steps=10,
monitor='val/loss',
*args,
**kwargs):
super().__init__(*args, **kwargs)
self.num_classes = num_classes
# get latest config of diffusion model
diffusion_config = natsorted(glob(os.path.join(diffusion_path, 'configs', '*-project.yaml')))[-1]
self.diffusion_config = OmegaConf.load(diffusion_config).model
self.diffusion_config.params.ckpt_path = diffusion_ckpt_path
self.load_diffusion()
self.monitor = monitor
self.numd = self.diffusion_model.first_stage_model.encoder.num_resolutions - 1
self.log_time_interval = self.diffusion_model.num_timesteps // log_steps
self.log_steps = log_steps
self.label_key = label_key if not hasattr(self.diffusion_model, 'cond_stage_key') \
else self.diffusion_model.cond_stage_key
assert self.label_key is not None, 'label_key neither in diffusion model nor in model.params'
if self.label_key not in __models__:
raise NotImplementedError()
self.load_classifier(ckpt_path, pool)
self.scheduler_config = scheduler_config
self.use_scheduler = self.scheduler_config is not None
self.weight_decay = weight_decay
def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
sd = torch.load(path, map_location="cpu")
if "state_dict" in list(sd.keys()):
sd = sd["state_dict"]
keys = list(sd.keys())
for k in keys:
for ik in ignore_keys:
if k.startswith(ik):
print("Deleting key {} from state_dict.".format(k))
del sd[k]
missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
sd, strict=False)
print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
if len(missing) > 0:
print(f"Missing Keys: {missing}")
if len(unexpected) > 0:
print(f"Unexpected Keys: {unexpected}")
def load_diffusion(self):
model = instantiate_from_config(self.diffusion_config)
self.diffusion_model = model.eval()
self.diffusion_model.train = disabled_train
for param in self.diffusion_model.parameters():
param.requires_grad = False
def load_classifier(self, ckpt_path, pool):
model_config = deepcopy(self.diffusion_config.params.unet_config.params)
model_config.in_channels = self.diffusion_config.params.unet_config.params.out_channels
model_config.out_channels = self.num_classes
if self.label_key == 'class_label':
model_config.pool = pool
self.model = __models__[self.label_key](**model_config)
if ckpt_path is not None:
print('#####################################################################')
print(f'load from ckpt "{ckpt_path}"')
print('#####################################################################')
self.init_from_ckpt(ckpt_path)
@torch.no_grad()
def get_x_noisy(self, x, t, noise=None):
noise = default(noise, lambda: torch.randn_like(x))
continuous_sqrt_alpha_cumprod = None
if self.diffusion_model.use_continuous_noise:
continuous_sqrt_alpha_cumprod = self.diffusion_model.sample_continuous_noise_level(x.shape[0], t + 1)
# todo: make sure t+1 is correct here
return self.diffusion_model.q_sample(x_start=x, t=t, noise=noise,
continuous_sqrt_alpha_cumprod=continuous_sqrt_alpha_cumprod)
def forward(self, x_noisy, t, *args, **kwargs):
return self.model(x_noisy, t)
@torch.no_grad()
def get_input(self, batch, k):
x = batch[k]
if len(x.shape) == 3:
x = x[..., None]
x = rearrange(x, 'b h w c -> b c h w')
x = x.to(memory_format=torch.contiguous_format).float()
return x
@torch.no_grad()
def get_conditioning(self, batch, k=None):
if k is None:
k = self.label_key
assert k is not None, 'Needs to provide label key'
targets = batch[k].to(self.device)
if self.label_key == 'segmentation':
targets = rearrange(targets, 'b h w c -> b c h w')
for down in range(self.numd):
h, w = targets.shape[-2:]
targets = F.interpolate(targets, size=(h // 2, w // 2), mode='nearest')
# targets = rearrange(targets,'b c h w -> b h w c')
return targets
def compute_top_k(self, logits, labels, k, reduction="mean"):
_, top_ks = torch.topk(logits, k, dim=1)
if reduction == "mean":
return (top_ks == labels[:, None]).float().sum(dim=-1).mean().item()
elif reduction == "none":
return (top_ks == labels[:, None]).float().sum(dim=-1)
def on_train_epoch_start(self):
# save some memory
self.diffusion_model.model.to('cpu')
@torch.no_grad()
def write_logs(self, loss, logits, targets):
log_prefix = 'train' if self.training else 'val'
log = {}
log[f"{log_prefix}/loss"] = loss.mean()
log[f"{log_prefix}/acc@1"] = self.compute_top_k(
logits, targets, k=1, reduction="mean"
)
log[f"{log_prefix}/acc@5"] = self.compute_top_k(
logits, targets, k=5, reduction="mean"
)
self.log_dict(log, prog_bar=False, logger=True, on_step=self.training, on_epoch=True)
self.log('loss', log[f"{log_prefix}/loss"], prog_bar=True, logger=False)
self.log('global_step', self.global_step, logger=False, on_epoch=False, prog_bar=True)
lr = self.optimizers().param_groups[0]['lr']
self.log('lr_abs', lr, on_step=True, logger=True, on_epoch=False, prog_bar=True)
def shared_step(self, batch, t=None):
x, *_ = self.diffusion_model.get_input(batch, k=self.diffusion_model.first_stage_key)
targets = self.get_conditioning(batch)
if targets.dim() == 4:
targets = targets.argmax(dim=1)
if t is None:
t = torch.randint(0, self.diffusion_model.num_timesteps, (x.shape[0],), device=self.device).long()
else:
t = torch.full(size=(x.shape[0],), fill_value=t, device=self.device).long()
x_noisy = self.get_x_noisy(x, t)
logits = self(x_noisy, t)
loss = F.cross_entropy(logits, targets, reduction='none')
self.write_logs(loss.detach(), logits.detach(), targets.detach())
loss = loss.mean()
return loss, logits, x_noisy, targets
def training_step(self, batch, batch_idx):
loss, *_ = self.shared_step(batch)
return loss
def reset_noise_accs(self):
self.noisy_acc = {t: {'acc@1': [], 'acc@5': []} for t in
range(0, self.diffusion_model.num_timesteps, self.diffusion_model.log_every_t)}
def on_validation_start(self):
self.reset_noise_accs()
@torch.no_grad()
def validation_step(self, batch, batch_idx):
loss, *_ = self.shared_step(batch)
for t in self.noisy_acc:
_, logits, _, targets = self.shared_step(batch, t)
self.noisy_acc[t]['acc@1'].append(self.compute_top_k(logits, targets, k=1, reduction='mean'))
self.noisy_acc[t]['acc@5'].append(self.compute_top_k(logits, targets, k=5, reduction='mean'))
return loss
def configure_optimizers(self):
optimizer = AdamW(self.model.parameters(), lr=self.learning_rate, weight_decay=self.weight_decay)
if self.use_scheduler:
scheduler = instantiate_from_config(self.scheduler_config)
print("Setting up LambdaLR scheduler...")
scheduler = [
{
'scheduler': LambdaLR(optimizer, lr_lambda=scheduler.schedule),
'interval': 'step',
'frequency': 1
}]
return [optimizer], scheduler
return optimizer
@torch.no_grad()
def log_images(self, batch, N=8, *args, **kwargs):
log = dict()
x = self.get_input(batch, self.diffusion_model.first_stage_key)
log['inputs'] = x
y = self.get_conditioning(batch)
if self.label_key == 'class_label':
y = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
log['labels'] = y
if ismap(y):
log['labels'] = self.diffusion_model.to_rgb(y)
for step in range(self.log_steps):
current_time = step * self.log_time_interval
_, logits, x_noisy, _ = self.shared_step(batch, t=current_time)
log[f'inputs@t{current_time}'] = x_noisy
pred = F.one_hot(logits.argmax(dim=1), num_classes=self.num_classes)
pred = rearrange(pred, 'b h w c -> b c h w')
log[f'pred@t{current_time}'] = self.diffusion_model.to_rgb(pred)
for key in log:
log[key] = log[key][:N]
return log

248
ldm/models/diffusion/ddim.py Normal file
View File

@ -0,0 +1,248 @@
"""SAMPLING ONLY."""
import torch
import numpy as np
from tqdm import tqdm
from functools import partial
from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, \
extract_into_tensor
class DDIMSampler(object):
def __init__(self, model, schedule="linear", **kwargs):
super().__init__()
self.model = model
self.ddpm_num_timesteps = model.num_timesteps
self.schedule = schedule
def register_buffer(self, name, attr):
if type(attr) == torch.Tensor:
if attr.device != torch.device("cuda"):
attr = attr.to(torch.device("cuda"))
setattr(self, name, attr)
def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
alphas_cumprod = self.model.alphas_cumprod
assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
self.register_buffer('betas', to_torch(self.model.betas))
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
# calculations for diffusion q(x_t | x_{t-1}) and others
self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
# ddim sampling parameters
ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
ddim_timesteps=self.ddim_timesteps,
eta=ddim_eta,verbose=verbose)
self.register_buffer('ddim_sigmas', ddim_sigmas)
self.register_buffer('ddim_alphas', ddim_alphas)
self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
(1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
1 - self.alphas_cumprod / self.alphas_cumprod_prev))
self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
@torch.no_grad()
def sample(self,
S,
batch_size,
shape,
conditioning=None,
callback=None,
normals_sequence=None,
img_callback=None,
quantize_x0=False,
eta=0.,
mask=None,
x0=None,
temperature=1.,
noise_dropout=0.,
score_corrector=None,
corrector_kwargs=None,
verbose=True,
x_T=None,
log_every_t=100,
unconditional_guidance_scale=1.,
unconditional_conditioning=None,
# this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
**kwargs
):
if conditioning is not None:
if isinstance(conditioning, dict):
cbs = conditioning[list(conditioning.keys())[0]].shape[0]
if cbs != batch_size:
print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
else:
if conditioning.shape[0] != batch_size:
print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
# sampling
C, H, W = shape
size = (batch_size, C, H, W)
print(f'Data shape for DDIM sampling is {size}, eta {eta}')
samples, intermediates = self.ddim_sampling(conditioning, size,
callback=callback,
img_callback=img_callback,
quantize_denoised=quantize_x0,
mask=mask, x0=x0,
ddim_use_original_steps=False,
noise_dropout=noise_dropout,
temperature=temperature,
score_corrector=score_corrector,
corrector_kwargs=corrector_kwargs,
x_T=x_T,
log_every_t=log_every_t,
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=unconditional_conditioning,
)
return samples, intermediates
@torch.no_grad()
def ddim_sampling(self, cond, shape,
x_T=None, ddim_use_original_steps=False,
callback=None, timesteps=None, quantize_denoised=False,
mask=None, x0=None, img_callback=None, log_every_t=100,
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
unconditional_guidance_scale=1., unconditional_conditioning=None,):
device = self.model.betas.device
b = shape[0]
if x_T is None:
img = torch.randn(shape, device=device)
else:
img = x_T
if timesteps is None:
timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
elif timesteps is not None and not ddim_use_original_steps:
subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
timesteps = self.ddim_timesteps[:subset_end]
intermediates = {'x_inter': [img], 'pred_x0': [img]}
time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
print(f"Running DDIM Sampling with {total_steps} timesteps")
iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
for i, step in enumerate(iterator):
index = total_steps - i - 1
ts = torch.full((b,), step, device=device, dtype=torch.long)
if mask is not None:
assert x0 is not None
img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass?
img = img_orig * mask + (1. - mask) * img
outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
quantize_denoised=quantize_denoised, temperature=temperature,
noise_dropout=noise_dropout, score_corrector=score_corrector,
corrector_kwargs=corrector_kwargs,
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=unconditional_conditioning)
img, pred_x0 = outs
if callback: callback(i)
if img_callback: img_callback(pred_x0, i)
if index % log_every_t == 0 or index == total_steps - 1:
intermediates['x_inter'].append(img)
intermediates['pred_x0'].append(pred_x0)
return img, intermediates
@torch.no_grad()
def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
unconditional_guidance_scale=1., unconditional_conditioning=None):
b, *_, device = *x.shape, x.device
if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
e_t = self.model.apply_model(x, t, c)
else:
x_in = torch.cat([x] * 2)
t_in = torch.cat([t] * 2)
c_in = torch.cat([unconditional_conditioning, c])
e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
if score_corrector is not None:
assert self.model.parameterization == "eps"
e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
# select parameters corresponding to the currently considered timestep
a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
# current prediction for x_0
pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
if quantize_denoised:
pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
# direction pointing to x_t
dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
if noise_dropout > 0.:
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
return x_prev, pred_x0
@torch.no_grad()
def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
# fast, but does not allow for exact reconstruction
# t serves as an index to gather the correct alphas
if use_original_steps:
sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
else:
sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
if noise is None:
noise = torch.randn_like(x0)
return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)
@torch.no_grad()
def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
use_original_steps=False, z_mask = None, x0=None):
timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
timesteps = timesteps[:t_start]
time_range = np.flip(timesteps)
total_steps = timesteps.shape[0]
print(f"Running DDIM Sampling with {total_steps} timesteps")
iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
x_dec = x_latent
for i, step in enumerate(iterator):
index = total_steps - i - 1
ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
if z_mask is not None and i < total_steps - 2:
assert x0 is not None
img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass?
mask_inv = 1. - z_mask
x_dec = (img_orig * mask_inv) + (z_mask * x_dec)
x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=unconditional_conditioning)
return x_dec

1445
ldm/models/diffusion/ddpm.py Normal file
View File

File diff suppressed because it is too large Load Diff

236
ldm/models/diffusion/plms.py Normal file
View File

@ -0,0 +1,236 @@
"""SAMPLING ONLY."""
import torch
import numpy as np
from tqdm import tqdm
from functools import partial
from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like
class PLMSSampler(object):
def __init__(self, model, schedule="linear", **kwargs):
super().__init__()
self.model = model
self.ddpm_num_timesteps = model.num_timesteps
self.schedule = schedule
def register_buffer(self, name, attr):
if type(attr) == torch.Tensor:
if attr.device != torch.device("cuda"):
attr = attr.to(torch.device("cuda"))
setattr(self, name, attr)
def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
if ddim_eta != 0:
raise ValueError('ddim_eta must be 0 for PLMS')
self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
alphas_cumprod = self.model.alphas_cumprod
assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
self.register_buffer('betas', to_torch(self.model.betas))
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
# calculations for diffusion q(x_t | x_{t-1}) and others
self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
# ddim sampling parameters
ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
ddim_timesteps=self.ddim_timesteps,
eta=ddim_eta,verbose=verbose)
self.register_buffer('ddim_sigmas', ddim_sigmas)
self.register_buffer('ddim_alphas', ddim_alphas)
self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
(1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
1 - self.alphas_cumprod / self.alphas_cumprod_prev))
self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
@torch.no_grad()
def sample(self,
S,
batch_size,
shape,
conditioning=None,
callback=None,
normals_sequence=None,
img_callback=None,
quantize_x0=False,
eta=0.,
mask=None,
x0=None,
temperature=1.,
noise_dropout=0.,
score_corrector=None,
corrector_kwargs=None,
verbose=True,
x_T=None,
log_every_t=100,
unconditional_guidance_scale=1.,
unconditional_conditioning=None,
# this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
**kwargs
):
if conditioning is not None:
if isinstance(conditioning, dict):
cbs = conditioning[list(conditioning.keys())[0]].shape[0]
if cbs != batch_size:
print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
else:
if conditioning.shape[0] != batch_size:
print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
# sampling
C, H, W = shape
size = (batch_size, C, H, W)
print(f'Data shape for PLMS sampling is {size}')
samples, intermediates = self.plms_sampling(conditioning, size,
callback=callback,
img_callback=img_callback,
quantize_denoised=quantize_x0,
mask=mask, x0=x0,
ddim_use_original_steps=False,
noise_dropout=noise_dropout,
temperature=temperature,
score_corrector=score_corrector,
corrector_kwargs=corrector_kwargs,
x_T=x_T,
log_every_t=log_every_t,
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=unconditional_conditioning,
)
return samples, intermediates
@torch.no_grad()
def plms_sampling(self, cond, shape,
x_T=None, ddim_use_original_steps=False,
callback=None, timesteps=None, quantize_denoised=False,
mask=None, x0=None, img_callback=None, log_every_t=100,
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
unconditional_guidance_scale=1., unconditional_conditioning=None,):
device = self.model.betas.device
b = shape[0]
if x_T is None:
img = torch.randn(shape, device=device)
else:
img = x_T
if timesteps is None:
timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
elif timesteps is not None and not ddim_use_original_steps:
subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
timesteps = self.ddim_timesteps[:subset_end]
intermediates = {'x_inter': [img], 'pred_x0': [img]}
time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps)
total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
print(f"Running PLMS Sampling with {total_steps} timesteps")
iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps)
old_eps = []
for i, step in enumerate(iterator):
index = total_steps - i - 1
ts = torch.full((b,), step, device=device, dtype=torch.long)
ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long)
if mask is not None:
assert x0 is not None
img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass?
img = img_orig * mask + (1. - mask) * img
outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
quantize_denoised=quantize_denoised, temperature=temperature,
noise_dropout=noise_dropout, score_corrector=score_corrector,
corrector_kwargs=corrector_kwargs,
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=unconditional_conditioning,
old_eps=old_eps, t_next=ts_next)
img, pred_x0, e_t = outs
old_eps.append(e_t)
if len(old_eps) >= 4:
old_eps.pop(0)
if callback: callback(i)
if img_callback: img_callback(pred_x0, i)
if index % log_every_t == 0 or index == total_steps - 1:
intermediates['x_inter'].append(img)
intermediates['pred_x0'].append(pred_x0)
return img, intermediates
@torch.no_grad()
def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None):
b, *_, device = *x.shape, x.device
def get_model_output(x, t):
if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
e_t = self.model.apply_model(x, t, c)
else:
x_in = torch.cat([x] * 2)
t_in = torch.cat([t] * 2)
c_in = torch.cat([unconditional_conditioning, c])
e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
if score_corrector is not None:
assert self.model.parameterization == "eps"
e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
return e_t
alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
def get_x_prev_and_pred_x0(e_t, index):
# select parameters corresponding to the currently considered timestep
a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
# current prediction for x_0
pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
if quantize_denoised:
pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
# direction pointing to x_t
dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
if noise_dropout > 0.:
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
return x_prev, pred_x0
e_t = get_model_output(x, t)
if len(old_eps) == 0:
# Pseudo Improved Euler (2nd order)
x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
e_t_next = get_model_output(x_prev, t_next)
e_t_prime = (e_t + e_t_next) / 2
elif len(old_eps) == 1:
# 2nd order Pseudo Linear Multistep (Adams-Bashforth)
e_t_prime = (3 * e_t - old_eps[-1]) / 2
elif len(old_eps) == 2:
# 3nd order Pseudo Linear Multistep (Adams-Bashforth)
e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
elif len(old_eps) >= 3:
# 4nd order Pseudo Linear Multistep (Adams-Bashforth)
e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24
x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
return x_prev, pred_x0, e_t

261
ldm/modules/attention.py Normal file
View File

@ -0,0 +1,261 @@
from inspect import isfunction
import math
import torch
import torch.nn.functional as F
from torch import nn, einsum
from einops import rearrange, repeat
from ldm.modules.diffusionmodules.util import checkpoint
def exists(val):
return val is not None
def uniq(arr):
return{el: True for el in arr}.keys()
def default(val, d):
if exists(val):
return val
return d() if isfunction(d) else d
def max_neg_value(t):
return -torch.finfo(t.dtype).max
def init_(tensor):
dim = tensor.shape[-1]
std = 1 / math.sqrt(dim)
tensor.uniform_(-std, std)
return tensor
# feedforward
class GEGLU(nn.Module):
def __init__(self, dim_in, dim_out):
super().__init__()
self.proj = nn.Linear(dim_in, dim_out * 2)
def forward(self, x):
x, gate = self.proj(x).chunk(2, dim=-1)
return x * F.gelu(gate)
class FeedForward(nn.Module):
def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
super().__init__()
inner_dim = int(dim * mult)
dim_out = default(dim_out, dim)
project_in = nn.Sequential(
nn.Linear(dim, inner_dim),
nn.GELU()
) if not glu else GEGLU(dim, inner_dim)
self.net = nn.Sequential(
project_in,
nn.Dropout(dropout),
nn.Linear(inner_dim, dim_out)
)
def forward(self, x):
return self.net(x)
def zero_module(module):
"""
Zero out the parameters of a module and return it.
"""
for p in module.parameters():
p.detach().zero_()
return module
def Normalize(in_channels):
return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
class LinearAttention(nn.Module):
def __init__(self, dim, heads=4, dim_head=32):
super().__init__()
self.heads = heads
hidden_dim = dim_head * heads
self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
self.to_out = nn.Conv2d(hidden_dim, dim, 1)
def forward(self, x):
b, c, h, w = x.shape
qkv = self.to_qkv(x)
q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
k = k.softmax(dim=-1)
context = torch.einsum('bhdn,bhen->bhde', k, v)
out = torch.einsum('bhde,bhdn->bhen', context, q)
out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
return self.to_out(out)
class SpatialSelfAttention(nn.Module):
def __init__(self, in_channels):
super().__init__()
self.in_channels = in_channels
self.norm = Normalize(in_channels)
self.q = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0)
self.k = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0)
self.v = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0)
self.proj_out = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0)
def forward(self, x):
h_ = x
h_ = self.norm(h_)
q = self.q(h_)
k = self.k(h_)
v = self.v(h_)
# compute attention
b,c,h,w = q.shape
q = rearrange(q, 'b c h w -> b (h w) c')
k = rearrange(k, 'b c h w -> b c (h w)')
w_ = torch.einsum('bij,bjk->bik', q, k)
w_ = w_ * (int(c)**(-0.5))
w_ = torch.nn.functional.softmax(w_, dim=2)
# attend to values
v = rearrange(v, 'b c h w -> b c (h w)')
w_ = rearrange(w_, 'b i j -> b j i')
h_ = torch.einsum('bij,bjk->bik', v, w_)
h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
h_ = self.proj_out(h_)
return x+h_
class CrossAttention(nn.Module):
def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.):
super().__init__()
inner_dim = dim_head * heads
context_dim = default(context_dim, query_dim)
self.scale = dim_head ** -0.5
self.heads = heads
self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
self.to_out = nn.Sequential(
nn.Linear(inner_dim, query_dim),
nn.Dropout(dropout)
)
def forward(self, x, context=None, mask=None):
h = self.heads
q = self.to_q(x)
context = default(context, x)
k = self.to_k(context)
v = self.to_v(context)
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
if exists(mask):
mask = rearrange(mask, 'b ... -> b (...)')
max_neg_value = -torch.finfo(sim.dtype).max
mask = repeat(mask, 'b j -> (b h) () j', h=h)
sim.masked_fill_(~mask, max_neg_value)
# attention, what we cannot get enough of
attn = sim.softmax(dim=-1)
out = einsum('b i j, b j d -> b i d', attn, v)
out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
return self.to_out(out)
class BasicTransformerBlock(nn.Module):
def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True):
super().__init__()
self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout) # is a self-attention
self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim,
heads=n_heads, dim_head=d_head, dropout=dropout) # is self-attn if context is none
self.norm1 = nn.LayerNorm(dim)
self.norm2 = nn.LayerNorm(dim)
self.norm3 = nn.LayerNorm(dim)
self.checkpoint = checkpoint
def forward(self, x, context=None):
return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
def _forward(self, x, context=None):
x = self.attn1(self.norm1(x)) + x
x = self.attn2(self.norm2(x), context=context) + x
x = self.ff(self.norm3(x)) + x
return x
class SpatialTransformer(nn.Module):
"""
Transformer block for image-like data.
First, project the input (aka embedding)
and reshape to b, t, d.
Then apply standard transformer action.
Finally, reshape to image
"""
def __init__(self, in_channels, n_heads, d_head,
depth=1, dropout=0., context_dim=None):
super().__init__()
self.in_channels = in_channels
inner_dim = n_heads * d_head
self.norm = Normalize(in_channels)
self.proj_in = nn.Conv2d(in_channels,
inner_dim,
kernel_size=1,
stride=1,
padding=0)
self.transformer_blocks = nn.ModuleList(
[BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim)
for d in range(depth)]
)
self.proj_out = zero_module(nn.Conv2d(inner_dim,
in_channels,
kernel_size=1,
stride=1,
padding=0))
def forward(self, x, context=None):
# note: if no context is given, cross-attention defaults to self-attention
b, c, h, w = x.shape
x_in = x
x = self.norm(x)
x = self.proj_in(x)
x = rearrange(x, 'b c h w -> b (h w) c')
for block in self.transformer_blocks:
x = block(x, context=context)
x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w)
x = self.proj_out(x)
return x + x_in

0
ldm/modules/diffusionmodules/__init__.py Normal file
View File

835
ldm/modules/diffusionmodules/model.py Normal file
View File

@ -0,0 +1,835 @@
# pytorch_diffusion + derived encoder decoder
import math
import torch
import torch.nn as nn
import numpy as np
from einops import rearrange
from ldm.util import instantiate_from_config
from ldm.modules.attention import LinearAttention
def get_timestep_embedding(timesteps, embedding_dim):
"""
This matches the implementation in Denoising Diffusion Probabilistic Models:
From Fairseq.
Build sinusoidal embeddings.
This matches the implementation in tensor2tensor, but differs slightly
from the description in Section 3.5 of "Attention Is All You Need".
"""
assert len(timesteps.shape) == 1
half_dim = embedding_dim // 2
emb = math.log(10000) / (half_dim - 1)
emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
emb = emb.to(device=timesteps.device)
emb = timesteps.float()[:, None] * emb[None, :]
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
if embedding_dim % 2 == 1: # zero pad
emb = torch.nn.functional.pad(emb, (0,1,0,0))
return emb
def nonlinearity(x):
# swish
return x*torch.sigmoid(x)
def Normalize(in_channels, num_groups=32):
return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
class Upsample(nn.Module):
def __init__(self, in_channels, with_conv):
super().__init__()
self.with_conv = with_conv
if self.with_conv:
self.conv = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=3,
stride=1,
padding=1)
def forward(self, x):
x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
if self.with_conv:
x = self.conv(x)
return x
class Downsample(nn.Module):
def __init__(self, in_channels, with_conv):
super().__init__()
self.with_conv = with_conv
if self.with_conv:
# no asymmetric padding in torch conv, must do it ourselves
self.conv = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=3,
stride=2,
padding=0)
def forward(self, x):
if self.with_conv:
pad = (0,1,0,1)
x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
x = self.conv(x)
else:
x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
return x
class ResnetBlock(nn.Module):
def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
dropout, temb_channels=512):
super().__init__()
self.in_channels = in_channels
out_channels = in_channels if out_channels is None else out_channels
self.out_channels = out_channels
self.use_conv_shortcut = conv_shortcut
self.norm1 = Normalize(in_channels)
self.conv1 = torch.nn.Conv2d(in_channels,
out_channels,
kernel_size=3,
stride=1,
padding=1)
if temb_channels > 0:
self.temb_proj = torch.nn.Linear(temb_channels,
out_channels)
self.norm2 = Normalize(out_channels)
self.dropout = torch.nn.Dropout(dropout)
self.conv2 = torch.nn.Conv2d(out_channels,
out_channels,
kernel_size=3,
stride=1,
padding=1)
if self.in_channels != self.out_channels:
if self.use_conv_shortcut:
self.conv_shortcut = torch.nn.Conv2d(in_channels,
out_channels,
kernel_size=3,
stride=1,
padding=1)
else:
self.nin_shortcut = torch.nn.Conv2d(in_channels,
out_channels,
kernel_size=1,
stride=1,
padding=0)
def forward(self, x, temb):
h = x
h = self.norm1(h)
h = nonlinearity(h)
h = self.conv1(h)
if temb is not None:
h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None]
h = self.norm2(h)
h = nonlinearity(h)
h = self.dropout(h)
h = self.conv2(h)
if self.in_channels != self.out_channels:
if self.use_conv_shortcut:
x = self.conv_shortcut(x)
else:
x = self.nin_shortcut(x)
return x+h
class LinAttnBlock(LinearAttention):
"""to match AttnBlock usage"""
def __init__(self, in_channels):
super().__init__(dim=in_channels, heads=1, dim_head=in_channels)
class AttnBlock(nn.Module):
def __init__(self, in_channels):
super().__init__()
self.in_channels = in_channels
self.norm = Normalize(in_channels)
self.q = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0)
self.k = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0)
self.v = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0)
self.proj_out = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0)
def forward(self, x):
h_ = x
h_ = self.norm(h_)
q = self.q(h_)
k = self.k(h_)
v = self.v(h_)
# compute attention
b,c,h,w = q.shape
q = q.reshape(b,c,h*w)
q = q.permute(0,2,1) # b,hw,c
k = k.reshape(b,c,h*w) # b,c,hw
w_ = torch.bmm(q,k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
w_ = w_ * (int(c)**(-0.5))
w_ = torch.nn.functional.softmax(w_, dim=2)
# attend to values
v = v.reshape(b,c,h*w)
w_ = w_.permute(0,2,1) # b,hw,hw (first hw of k, second of q)
h_ = torch.bmm(v,w_) # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
h_ = h_.reshape(b,c,h,w)
h_ = self.proj_out(h_)
return x+h_
def make_attn(in_channels, attn_type="vanilla"):
assert attn_type in ["vanilla", "linear", "none"], f'attn_type {attn_type} unknown'
print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
if attn_type == "vanilla":
return AttnBlock(in_channels)
elif attn_type == "none":
return nn.Identity(in_channels)
else:
return LinAttnBlock(in_channels)
class Model(nn.Module):
def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
super().__init__()
if use_linear_attn: attn_type = "linear"
self.ch = ch
self.temb_ch = self.ch*4
self.num_resolutions = len(ch_mult)
self.num_res_blocks = num_res_blocks
self.resolution = resolution
self.in_channels = in_channels
self.use_timestep = use_timestep
if self.use_timestep:
# timestep embedding
self.temb = nn.Module()
self.temb.dense = nn.ModuleList([
torch.nn.Linear(self.ch,
self.temb_ch),
torch.nn.Linear(self.temb_ch,
self.temb_ch),
])
# downsampling
self.conv_in = torch.nn.Conv2d(in_channels,
self.ch,
kernel_size=3,
stride=1,
padding=1)
curr_res = resolution
in_ch_mult = (1,)+tuple(ch_mult)
self.down = nn.ModuleList()
for i_level in range(self.num_resolutions):
block = nn.ModuleList()
attn = nn.ModuleList()
block_in = ch*in_ch_mult[i_level]
block_out = ch*ch_mult[i_level]
for i_block in range(self.num_res_blocks):
block.append(ResnetBlock(in_channels=block_in,
out_channels=block_out,
temb_channels=self.temb_ch,
dropout=dropout))
block_in = block_out
if curr_res in attn_resolutions:
attn.append(make_attn(block_in, attn_type=attn_type))
down = nn.Module()
down.block = block
down.attn = attn
if i_level != self.num_resolutions-1:
down.downsample = Downsample(block_in, resamp_with_conv)
curr_res = curr_res // 2
self.down.append(down)
# middle
self.mid = nn.Module()
self.mid.block_1 = ResnetBlock(in_channels=block_in,
out_channels=block_in,
temb_channels=self.temb_ch,
dropout=dropout)
self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
self.mid.block_2 = ResnetBlock(in_channels=block_in,
out_channels=block_in,
temb_channels=self.temb_ch,
dropout=dropout)
# upsampling
self.up = nn.ModuleList()
for i_level in reversed(range(self.num_resolutions)):
block = nn.ModuleList()
attn = nn.ModuleList()
block_out = ch*ch_mult[i_level]
skip_in = ch*ch_mult[i_level]
for i_block in range(self.num_res_blocks+1):
if i_block == self.num_res_blocks:
skip_in = ch*in_ch_mult[i_level]
block.append(ResnetBlock(in_channels=block_in+skip_in,
out_channels=block_out,
temb_channels=self.temb_ch,
dropout=dropout))
block_in = block_out
if curr_res in attn_resolutions:
attn.append(make_attn(block_in, attn_type=attn_type))
up = nn.Module()
up.block = block
up.attn = attn
if i_level != 0:
up.upsample = Upsample(block_in, resamp_with_conv)
curr_res = curr_res * 2
self.up.insert(0, up) # prepend to get consistent order
# end
self.norm_out = Normalize(block_in)
self.conv_out = torch.nn.Conv2d(block_in,
out_ch,
kernel_size=3,
stride=1,
padding=1)
def forward(self, x, t=None, context=None):
#assert x.shape[2] == x.shape[3] == self.resolution
if context is not None:
# assume aligned context, cat along channel axis
x = torch.cat((x, context), dim=1)
if self.use_timestep:
# timestep embedding
assert t is not None
temb = get_timestep_embedding(t, self.ch)
temb = self.temb.dense[0](temb)
temb = nonlinearity(temb)
temb = self.temb.dense[1](temb)
else:
temb = None
# downsampling
hs = [self.conv_in(x)]
for i_level in range(self.num_resolutions):
for i_block in range(self.num_res_blocks):
h = self.down[i_level].block[i_block](hs[-1], temb)
if len(self.down[i_level].attn) > 0:
h = self.down[i_level].attn[i_block](h)
hs.append(h)
if i_level != self.num_resolutions-1:
hs.append(self.down[i_level].downsample(hs[-1]))
# middle
h = hs[-1]
h = self.mid.block_1(h, temb)
h = self.mid.attn_1(h)
h = self.mid.block_2(h, temb)
# upsampling
for i_level in reversed(range(self.num_resolutions)):
for i_block in range(self.num_res_blocks+1):
h = self.up[i_level].block[i_block](
torch.cat([h, hs.pop()], dim=1), temb)
if len(self.up[i_level].attn) > 0:
h = self.up[i_level].attn[i_block](h)
if i_level != 0:
h = self.up[i_level].upsample(h)
# end
h = self.norm_out(h)
h = nonlinearity(h)
h = self.conv_out(h)
return h
def get_last_layer(self):
return self.conv_out.weight
class Encoder(nn.Module):
def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla",
**ignore_kwargs):
super().__init__()
if use_linear_attn: attn_type = "linear"
self.ch = ch
self.temb_ch = 0
self.num_resolutions = len(ch_mult)
self.num_res_blocks = num_res_blocks
self.resolution = resolution
self.in_channels = in_channels
# downsampling
self.conv_in = torch.nn.Conv2d(in_channels,
self.ch,
kernel_size=3,
stride=1,
padding=1)
curr_res = resolution
in_ch_mult = (1,)+tuple(ch_mult)
self.in_ch_mult = in_ch_mult
self.down = nn.ModuleList()
for i_level in range(self.num_resolutions):
block = nn.ModuleList()
attn = nn.ModuleList()
block_in = ch*in_ch_mult[i_level]
block_out = ch*ch_mult[i_level]
for i_block in range(self.num_res_blocks):
block.append(ResnetBlock(in_channels=block_in,
out_channels=block_out,
temb_channels=self.temb_ch,
dropout=dropout))
block_in = block_out
if curr_res in attn_resolutions:
attn.append(make_attn(block_in, attn_type=attn_type))
down = nn.Module()
down.block = block
down.attn = attn
if i_level != self.num_resolutions-1:
down.downsample = Downsample(block_in, resamp_with_conv)
curr_res = curr_res // 2
self.down.append(down)
# middle
self.mid = nn.Module()
self.mid.block_1 = ResnetBlock(in_channels=block_in,
out_channels=block_in,
temb_channels=self.temb_ch,
dropout=dropout)
self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
self.mid.block_2 = ResnetBlock(in_channels=block_in,
out_channels=block_in,
temb_channels=self.temb_ch,
dropout=dropout)
# end
self.norm_out = Normalize(block_in)
self.conv_out = torch.nn.Conv2d(block_in,
2*z_channels if double_z else z_channels,
kernel_size=3,
stride=1,
padding=1)
def forward(self, x):
# timestep embedding
temb = None
# downsampling
hs = [self.conv_in(x)]
for i_level in range(self.num_resolutions):
for i_block in range(self.num_res_blocks):
h = self.down[i_level].block[i_block](hs[-1], temb)
if len(self.down[i_level].attn) > 0:
h = self.down[i_level].attn[i_block](h)
hs.append(h)
if i_level != self.num_resolutions-1:
hs.append(self.down[i_level].downsample(hs[-1]))
# middle
h = hs[-1]
h = self.mid.block_1(h, temb)
h = self.mid.attn_1(h)
h = self.mid.block_2(h, temb)
# end
h = self.norm_out(h)
h = nonlinearity(h)
h = self.conv_out(h)
return h
class Decoder(nn.Module):
def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
attn_type="vanilla", **ignorekwargs):
super().__init__()
if use_linear_attn: attn_type = "linear"
self.ch = ch
self.temb_ch = 0
self.num_resolutions = len(ch_mult)
self.num_res_blocks = num_res_blocks
self.resolution = resolution
self.in_channels = in_channels
self.give_pre_end = give_pre_end
self.tanh_out = tanh_out
# compute in_ch_mult, block_in and curr_res at lowest res
in_ch_mult = (1,)+tuple(ch_mult)
block_in = ch*ch_mult[self.num_resolutions-1]
curr_res = resolution // 2**(self.num_resolutions-1)
self.z_shape = (1,z_channels,curr_res,curr_res)
print("Working with z of shape {} = {} dimensions.".format(
self.z_shape, np.prod(self.z_shape)))
# z to block_in
self.conv_in = torch.nn.Conv2d(z_channels,
block_in,
kernel_size=3,
stride=1,
padding=1)
# middle
self.mid = nn.Module()
self.mid.block_1 = ResnetBlock(in_channels=block_in,
out_channels=block_in,
temb_channels=self.temb_ch,
dropout=dropout)
self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
self.mid.block_2 = ResnetBlock(in_channels=block_in,
out_channels=block_in,
temb_channels=self.temb_ch,
dropout=dropout)
# upsampling
self.up = nn.ModuleList()
for i_level in reversed(range(self.num_resolutions)):
block = nn.ModuleList()
attn = nn.ModuleList()
block_out = ch*ch_mult[i_level]
for i_block in range(self.num_res_blocks+1):
block.append(ResnetBlock(in_channels=block_in,
out_channels=block_out,
temb_channels=self.temb_ch,
dropout=dropout))
block_in = block_out
if curr_res in attn_resolutions:
attn.append(make_attn(block_in, attn_type=attn_type))
up = nn.Module()
up.block = block
up.attn = attn
if i_level != 0:
up.upsample = Upsample(block_in, resamp_with_conv)
curr_res = curr_res * 2
self.up.insert(0, up) # prepend to get consistent order
# end
self.norm_out = Normalize(block_in)
self.conv_out = torch.nn.Conv2d(block_in,
out_ch,
kernel_size=3,
stride=1,
padding=1)
def forward(self, z):
#assert z.shape[1:] == self.z_shape[1:]
self.last_z_shape = z.shape
# timestep embedding
temb = None
# z to block_in
h = self.conv_in(z)
# middle
h = self.mid.block_1(h, temb)
h = self.mid.attn_1(h)
h = self.mid.block_2(h, temb)
# upsampling
for i_level in reversed(range(self.num_resolutions)):
for i_block in range(self.num_res_blocks+1):
h = self.up[i_level].block[i_block](h, temb)
if len(self.up[i_level].attn) > 0:
h = self.up[i_level].attn[i_block](h)
if i_level != 0:
h = self.up[i_level].upsample(h)
# end
if self.give_pre_end:
return h
h = self.norm_out(h)
h = nonlinearity(h)
h = self.conv_out(h)
if self.tanh_out:
h = torch.tanh(h)
return h
class SimpleDecoder(nn.Module):
def __init__(self, in_channels, out_channels, *args, **kwargs):
super().__init__()
self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1),
ResnetBlock(in_channels=in_channels,
out_channels=2 * in_channels,
temb_channels=0, dropout=0.0),
ResnetBlock(in_channels=2 * in_channels,
out_channels=4 * in_channels,
temb_channels=0, dropout=0.0),
ResnetBlock(in_channels=4 * in_channels,
out_channels=2 * in_channels,
temb_channels=0, dropout=0.0),
nn.Conv2d(2*in_channels, in_channels, 1),
Upsample(in_channels, with_conv=True)])
# end
self.norm_out = Normalize(in_channels)
self.conv_out = torch.nn.Conv2d(in_channels,
out_channels,
kernel_size=3,
stride=1,
padding=1)
def forward(self, x):
for i, layer in enumerate(self.model):
if i in [1,2,3]:
x = layer(x, None)
else:
x = layer(x)
h = self.norm_out(x)
h = nonlinearity(h)
x = self.conv_out(h)
return x
class UpsampleDecoder(nn.Module):
def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution,
ch_mult=(2,2), dropout=0.0):
super().__init__()
# upsampling
self.temb_ch = 0
self.num_resolutions = len(ch_mult)
self.num_res_blocks = num_res_blocks
block_in = in_channels
curr_res = resolution // 2 ** (self.num_resolutions - 1)
self.res_blocks = nn.ModuleList()
self.upsample_blocks = nn.ModuleList()
for i_level in range(self.num_resolutions):
res_block = []
block_out = ch * ch_mult[i_level]
for i_block in range(self.num_res_blocks + 1):
res_block.append(ResnetBlock(in_channels=block_in,
out_channels=block_out,
temb_channels=self.temb_ch,
dropout=dropout))
block_in = block_out
self.res_blocks.append(nn.ModuleList(res_block))
if i_level != self.num_resolutions - 1:
self.upsample_blocks.append(Upsample(block_in, True))
curr_res = curr_res * 2
# end
self.norm_out = Normalize(block_in)
self.conv_out = torch.nn.Conv2d(block_in,
out_channels,
kernel_size=3,
stride=1,
padding=1)
def forward(self, x):
# upsampling
h = x
for k, i_level in enumerate(range(self.num_resolutions)):
for i_block in range(self.num_res_blocks + 1):
h = self.res_blocks[i_level][i_block](h, None)
if i_level != self.num_resolutions - 1:
h = self.upsample_blocks[k](h)
h = self.norm_out(h)
h = nonlinearity(h)
h = self.conv_out(h)
return h
class LatentRescaler(nn.Module):
def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
super().__init__()
# residual block, interpolate, residual block
self.factor = factor
self.conv_in = nn.Conv2d(in_channels,
mid_channels,
kernel_size=3,
stride=1,
padding=1)
self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
out_channels=mid_channels,
temb_channels=0,
dropout=0.0) for _ in range(depth)])
self.attn = AttnBlock(mid_channels)
self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
out_channels=mid_channels,
temb_channels=0,
dropout=0.0) for _ in range(depth)])
self.conv_out = nn.Conv2d(mid_channels,
out_channels,
kernel_size=1,
)
def forward(self, x):
x = self.conv_in(x)
for block in self.res_block1:
x = block(x, None)
x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor))))
x = self.attn(x)
for block in self.res_block2:
x = block(x, None)
x = self.conv_out(x)
return x
class MergedRescaleEncoder(nn.Module):
def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks,
attn_resolutions, dropout=0.0, resamp_with_conv=True,
ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1):
super().__init__()
intermediate_chn = ch * ch_mult[-1]
self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult,
z_channels=intermediate_chn, double_z=False, resolution=resolution,
attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv,
out_ch=None)
self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn,
mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth)
def forward(self, x):
x = self.encoder(x)
x = self.rescaler(x)
return x
class MergedRescaleDecoder(nn.Module):
def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8),
dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1):
super().__init__()
tmp_chn = z_channels*ch_mult[-1]
self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout,
resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks,
ch_mult=ch_mult, resolution=resolution, ch=ch)
self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn,
out_channels=tmp_chn, depth=rescale_module_depth)
def forward(self, x):
x = self.rescaler(x)
x = self.decoder(x)
return x
class Upsampler(nn.Module):
def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
super().__init__()
assert out_size >= in_size
num_blocks = int(np.log2(out_size//in_size))+1
factor_up = 1.+ (out_size % in_size)
print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}")
self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels,
out_channels=in_channels)
self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2,
attn_resolutions=[], in_channels=None, ch=in_channels,
ch_mult=[ch_mult for _ in range(num_blocks)])
def forward(self, x):
x = self.rescaler(x)
x = self.decoder(x)
return x
class Resize(nn.Module):
def __init__(self, in_channels=None, learned=False, mode="bilinear"):
super().__init__()
self.with_conv = learned
self.mode = mode
if self.with_conv:
print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode")
raise NotImplementedError()
assert in_channels is not None
# no asymmetric padding in torch conv, must do it ourselves
self.conv = torch.nn.Conv2d(in_channels,
in_channels,
kernel_size=4,
stride=2,
padding=1)
def forward(self, x, scale_factor=1.0):
if scale_factor==1.0:
return x
else:
x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor)
return x
class FirstStagePostProcessor(nn.Module):
def __init__(self, ch_mult:list, in_channels,
pretrained_model:nn.Module=None,
reshape=False,
n_channels=None,
dropout=0.,
pretrained_config=None):
super().__init__()
if pretrained_config is None:
assert pretrained_model is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
self.pretrained_model = pretrained_model
else:
assert pretrained_config is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
self.instantiate_pretrained(pretrained_config)
self.do_reshape = reshape
if n_channels is None:
n_channels = self.pretrained_model.encoder.ch
self.proj_norm = Normalize(in_channels,num_groups=in_channels//2)
self.proj = nn.Conv2d(in_channels,n_channels,kernel_size=3,
stride=1,padding=1)
blocks = []
downs = []
ch_in = n_channels
for m in ch_mult:
blocks.append(ResnetBlock(in_channels=ch_in,out_channels=m*n_channels,dropout=dropout))
ch_in = m * n_channels
downs.append(Downsample(ch_in, with_conv=False))
self.model = nn.ModuleList(blocks)
self.downsampler = nn.ModuleList(downs)
def instantiate_pretrained(self, config):
model = instantiate_from_config(config)
self.pretrained_model = model.eval()
# self.pretrained_model.train = False
for param in self.pretrained_model.parameters():
param.requires_grad = False
@torch.no_grad()
def encode_with_pretrained(self,x):
c = self.pretrained_model.encode(x)
if isinstance(c, DiagonalGaussianDistribution):
c = c.mode()
return c
def forward(self,x):
z_fs = self.encode_with_pretrained(x)
z = self.proj_norm(z_fs)
z = self.proj(z)
z = nonlinearity(z)
for submodel, downmodel in zip(self.model,self.downsampler):
z = submodel(z,temb=None)
z = downmodel(z)
if self.do_reshape:
z = rearrange(z,'b c h w -> b (h w) c')
return z

961
ldm/modules/diffusionmodules/openaimodel.py Normal file
View File

@ -0,0 +1,961 @@
from abc import abstractmethod
from functools import partial
import math
from typing import Iterable
import numpy as np
import torch as th
import torch.nn as nn
import torch.nn.functional as F
from ldm.modules.diffusionmodules.util import (
checkpoint,
conv_nd,
linear,
avg_pool_nd,
zero_module,
normalization,
timestep_embedding,
)
from ldm.modules.attention import SpatialTransformer
# dummy replace
def convert_module_to_f16(x):
pass
def convert_module_to_f32(x):
pass
## go
class AttentionPool2d(nn.Module):
"""
Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
"""
def __init__(
self,
spacial_dim: int,
embed_dim: int,
num_heads_channels: int,
output_dim: int = None,
):
super().__init__()
self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5)
self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1)
self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1)
self.num_heads = embed_dim // num_heads_channels
self.attention = QKVAttention(self.num_heads)
def forward(self, x):
b, c, *_spatial = x.shape
x = x.reshape(b, c, -1) # NC(HW)
x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) # NC(HW+1)
x = x + self.positional_embedding[None, :, :].to(x.dtype) # NC(HW+1)
x = self.qkv_proj(x)
x = self.attention(x)
x = self.c_proj(x)
return x[:, :, 0]
class TimestepBlock(nn.Module):
"""
Any module where forward() takes timestep embeddings as a second argument.
"""
@abstractmethod
def forward(self, x, emb):
"""
Apply the module to `x` given `emb` timestep embeddings.
"""
class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
"""
A sequential module that passes timestep embeddings to the children that
support it as an extra input.
"""
def forward(self, x, emb, context=None):
for layer in self:
if isinstance(layer, TimestepBlock):
x = layer(x, emb)
elif isinstance(layer, SpatialTransformer):
x = layer(x, context)
else:
x = layer(x)
return x
class Upsample(nn.Module):
"""
An upsampling layer with an optional convolution.
:param channels: channels in the inputs and outputs.
:param use_conv: a bool determining if a convolution is applied.
:param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
upsampling occurs in the inner-two dimensions.
"""
def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
self.use_conv = use_conv
self.dims = dims
if use_conv:
self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
def forward(self, x):
assert x.shape[1] == self.channels
if self.dims == 3:
x = F.interpolate(
x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
)
else:
x = F.interpolate(x, scale_factor=2, mode="nearest")
if self.use_conv:
x = self.conv(x)
return x
class TransposedUpsample(nn.Module):
'Learned 2x upsampling without padding'
def __init__(self, channels, out_channels=None, ks=5):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2)
def forward(self,x):
return self.up(x)
class Downsample(nn.Module):
"""
A downsampling layer with an optional convolution.
:param channels: channels in the inputs and outputs.
:param use_conv: a bool determining if a convolution is applied.
:param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
downsampling occurs in the inner-two dimensions.
"""
def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
self.use_conv = use_conv
self.dims = dims
stride = 2 if dims != 3 else (1, 2, 2)
if use_conv:
self.op = conv_nd(
dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
)
else:
assert self.channels == self.out_channels
self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
def forward(self, x):
assert x.shape[1] == self.channels
return self.op(x)
class ResBlock(TimestepBlock):
"""
A residual block that can optionally change the number of channels.
:param channels: the number of input channels.
:param emb_channels: the number of timestep embedding channels.
:param dropout: the rate of dropout.
:param out_channels: if specified, the number of out channels.
:param use_conv: if True and out_channels is specified, use a spatial
convolution instead of a smaller 1x1 convolution to change the
channels in the skip connection.
:param dims: determines if the signal is 1D, 2D, or 3D.
:param use_checkpoint: if True, use gradient checkpointing on this module.
:param up: if True, use this block for upsampling.
:param down: if True, use this block for downsampling.
"""
def __init__(
self,
channels,
emb_channels,
dropout,
out_channels=None,
use_conv=False,
use_scale_shift_norm=False,
dims=2,
use_checkpoint=False,
up=False,
down=False,
):
super().__init__()
self.channels = channels
self.emb_channels = emb_channels
self.dropout = dropout
self.out_channels = out_channels or channels
self.use_conv = use_conv
self.use_checkpoint = use_checkpoint
self.use_scale_shift_norm = use_scale_shift_norm
self.in_layers = nn.Sequential(
normalization(channels),
nn.SiLU(),
conv_nd(dims, channels, self.out_channels, 3, padding=1),
)
self.updown = up or down
if up:
self.h_upd = Upsample(channels, False, dims)
self.x_upd = Upsample(channels, False, dims)
elif down:
self.h_upd = Downsample(channels, False, dims)
self.x_upd = Downsample(channels, False, dims)
else:
self.h_upd = self.x_upd = nn.Identity()
self.emb_layers = nn.Sequential(
nn.SiLU(),
linear(
emb_channels,
2 * self.out_channels if use_scale_shift_norm else self.out_channels,
),
)
self.out_layers = nn.Sequential(
normalization(self.out_channels),
nn.SiLU(),
nn.Dropout(p=dropout),
zero_module(
conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
),
)
if self.out_channels == channels:
self.skip_connection = nn.Identity()
elif use_conv:
self.skip_connection = conv_nd(
dims, channels, self.out_channels, 3, padding=1
)
else:
self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
def forward(self, x, emb):
"""
Apply the block to a Tensor, conditioned on a timestep embedding.
:param x: an [N x C x ...] Tensor of features.
:param emb: an [N x emb_channels] Tensor of timestep embeddings.
:return: an [N x C x ...] Tensor of outputs.
"""
return checkpoint(
self._forward, (x, emb), self.parameters(), self.use_checkpoint
)
def _forward(self, x, emb):
if self.updown:
in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
h = in_rest(x)
h = self.h_upd(h)
x = self.x_upd(x)
h = in_conv(h)
else:
h = self.in_layers(x)
emb_out = self.emb_layers(emb).type(h.dtype)
while len(emb_out.shape) < len(h.shape):
emb_out = emb_out[..., None]
if self.use_scale_shift_norm:
out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
scale, shift = th.chunk(emb_out, 2, dim=1)
h = out_norm(h) * (1 + scale) + shift
h = out_rest(h)
else:
h = h + emb_out
h = self.out_layers(h)
return self.skip_connection(x) + h
class AttentionBlock(nn.Module):
"""
An attention block that allows spatial positions to attend to each other.
Originally ported from here, but adapted to the N-d case.
https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
"""
def __init__(
self,
channels,
num_heads=1,
num_head_channels=-1,
use_checkpoint=False,
use_new_attention_order=False,
):
super().__init__()
self.channels = channels
if num_head_channels == -1:
self.num_heads = num_heads
else:
assert (
channels % num_head_channels == 0
), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
self.num_heads = channels // num_head_channels
self.use_checkpoint = use_checkpoint
self.norm = normalization(channels)
self.qkv = conv_nd(1, channels, channels * 3, 1)
if use_new_attention_order:
# split qkv before split heads
self.attention = QKVAttention(self.num_heads)
else:
# split heads before split qkv
self.attention = QKVAttentionLegacy(self.num_heads)
self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
def forward(self, x):
return checkpoint(self._forward, (x,), self.parameters(), True) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!!
#return pt_checkpoint(self._forward, x) # pytorch
def _forward(self, x):
b, c, *spatial = x.shape
x = x.reshape(b, c, -1)
qkv = self.qkv(self.norm(x))
h = self.attention(qkv)
h = self.proj_out(h)
return (x + h).reshape(b, c, *spatial)
def count_flops_attn(model, _x, y):
"""
A counter for the `thop` package to count the operations in an
attention operation.
Meant to be used like:
macs, params = thop.profile(
model,
inputs=(inputs, timestamps),
custom_ops={QKVAttention: QKVAttention.count_flops},
)
"""
b, c, *spatial = y[0].shape
num_spatial = int(np.prod(spatial))
# We perform two matmuls with the same number of ops.
# The first computes the weight matrix, the second computes
# the combination of the value vectors.
matmul_ops = 2 * b * (num_spatial ** 2) * c
model.total_ops += th.DoubleTensor([matmul_ops])
class QKVAttentionLegacy(nn.Module):
"""
A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
"""
def __init__(self, n_heads):
super().__init__()
self.n_heads = n_heads
def forward(self, qkv):
"""
Apply QKV attention.
:param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
:return: an [N x (H * C) x T] tensor after attention.
"""
bs, width, length = qkv.shape
assert width % (3 * self.n_heads) == 0
ch = width // (3 * self.n_heads)
q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
scale = 1 / math.sqrt(math.sqrt(ch))
weight = th.einsum(
"bct,bcs->bts", q * scale, k * scale
) # More stable with f16 than dividing afterwards
weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
a = th.einsum("bts,bcs->bct", weight, v)
return a.reshape(bs, -1, length)
@staticmethod
def count_flops(model, _x, y):
return count_flops_attn(model, _x, y)
class QKVAttention(nn.Module):
"""
A module which performs QKV attention and splits in a different order.
"""
def __init__(self, n_heads):
super().__init__()
self.n_heads = n_heads
def forward(self, qkv):
"""
Apply QKV attention.
:param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
:return: an [N x (H * C) x T] tensor after attention.
"""
bs, width, length = qkv.shape
assert width % (3 * self.n_heads) == 0
ch = width // (3 * self.n_heads)
q, k, v = qkv.chunk(3, dim=1)
scale = 1 / math.sqrt(math.sqrt(ch))
weight = th.einsum(
"bct,bcs->bts",
(q * scale).view(bs * self.n_heads, ch, length),
(k * scale).view(bs * self.n_heads, ch, length),
) # More stable with f16 than dividing afterwards
weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length))
return a.reshape(bs, -1, length)
@staticmethod
def count_flops(model, _x, y):
return count_flops_attn(model, _x, y)
class UNetModel(nn.Module):
"""
The full UNet model with attention and timestep embedding.
:param in_channels: channels in the input Tensor.
:param model_channels: base channel count for the model.
:param out_channels: channels in the output Tensor.
:param num_res_blocks: number of residual blocks per downsample.
:param attention_resolutions: a collection of downsample rates at which
attention will take place. May be a set, list, or tuple.
For example, if this contains 4, then at 4x downsampling, attention
will be used.
:param dropout: the dropout probability.
:param channel_mult: channel multiplier for each level of the UNet.
:param conv_resample: if True, use learned convolutions for upsampling and
downsampling.
:param dims: determines if the signal is 1D, 2D, or 3D.
:param num_classes: if specified (as an int), then this model will be
class-conditional with `num_classes` classes.
:param use_checkpoint: use gradient checkpointing to reduce memory usage.
:param num_heads: the number of attention heads in each attention layer.
:param num_heads_channels: if specified, ignore num_heads and instead use
a fixed channel width per attention head.
:param num_heads_upsample: works with num_heads to set a different number
of heads for upsampling. Deprecated.
:param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
:param resblock_updown: use residual blocks for up/downsampling.
:param use_new_attention_order: use a different attention pattern for potentially
increased efficiency.
"""
def __init__(
self,
image_size,
in_channels,
model_channels,
out_channels,
num_res_blocks,
attention_resolutions,
dropout=0,
channel_mult=(1, 2, 4, 8),
conv_resample=True,
dims=2,
num_classes=None,
use_checkpoint=False,
use_fp16=False,
num_heads=-1,
num_head_channels=-1,
num_heads_upsample=-1,
use_scale_shift_norm=False,
resblock_updown=False,
use_new_attention_order=False,
use_spatial_transformer=False, # custom transformer support
transformer_depth=1, # custom transformer support
context_dim=None, # custom transformer support
n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model
legacy=True,
):
super().__init__()
if use_spatial_transformer:
assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
if context_dim is not None:
assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
from omegaconf.listconfig import ListConfig
if type(context_dim) == ListConfig:
context_dim = list(context_dim)
if num_heads_upsample == -1:
num_heads_upsample = num_heads
if num_heads == -1:
assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
if num_head_channels == -1:
assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
self.image_size = image_size
self.in_channels = in_channels
self.model_channels = model_channels
self.out_channels = out_channels
self.num_res_blocks = num_res_blocks
self.attention_resolutions = attention_resolutions
self.dropout = dropout
self.channel_mult = channel_mult
self.conv_resample = conv_resample
self.num_classes = num_classes
self.use_checkpoint = use_checkpoint
self.dtype = th.float16 if use_fp16 else th.float32
self.num_heads = num_heads
self.num_head_channels = num_head_channels
self.num_heads_upsample = num_heads_upsample
self.predict_codebook_ids = n_embed is not None
time_embed_dim = model_channels * 4
self.time_embed = nn.Sequential(
linear(model_channels, time_embed_dim),
nn.SiLU(),
linear(time_embed_dim, time_embed_dim),
)
if self.num_classes is not None:
self.label_emb = nn.Embedding(num_classes, time_embed_dim)
self.input_blocks = nn.ModuleList(
[
TimestepEmbedSequential(
conv_nd(dims, in_channels, model_channels, 3, padding=1)
)
]
)
self._feature_size = model_channels
input_block_chans = [model_channels]
ch = model_channels
ds = 1
for level, mult in enumerate(channel_mult):
for _ in range(num_res_blocks):
layers = [
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=mult * model_channels,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
)
]
ch = mult * model_channels
if ds in attention_resolutions:
if num_head_channels == -1:
dim_head = ch // num_heads
else:
num_heads = ch // num_head_channels
dim_head = num_head_channels
if legacy:
#num_heads = 1
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
layers.append(
AttentionBlock(
ch,
use_checkpoint=use_checkpoint,
num_heads=num_heads,
num_head_channels=dim_head,
use_new_attention_order=use_new_attention_order,
) if not use_spatial_transformer else SpatialTransformer(
ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
)
)
self.input_blocks.append(TimestepEmbedSequential(*layers))
self._feature_size += ch
input_block_chans.append(ch)
if level != len(channel_mult) - 1:
out_ch = ch
self.input_blocks.append(
TimestepEmbedSequential(
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=out_ch,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
down=True,
)
if resblock_updown
else Downsample(
ch, conv_resample, dims=dims, out_channels=out_ch
)
)
)
ch = out_ch
input_block_chans.append(ch)
ds *= 2
self._feature_size += ch
if num_head_channels == -1:
dim_head = ch // num_heads
else:
num_heads = ch // num_head_channels
dim_head = num_head_channels
if legacy:
#num_heads = 1
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
self.middle_block = TimestepEmbedSequential(
ResBlock(
ch,
time_embed_dim,
dropout,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
),
AttentionBlock(
ch,
use_checkpoint=use_checkpoint,
num_heads=num_heads,
num_head_channels=dim_head,
use_new_attention_order=use_new_attention_order,
) if not use_spatial_transformer else SpatialTransformer(
ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
),
ResBlock(
ch,
time_embed_dim,
dropout,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
),
)
self._feature_size += ch
self.output_blocks = nn.ModuleList([])
for level, mult in list(enumerate(channel_mult))[::-1]:
for i in range(num_res_blocks + 1):
ich = input_block_chans.pop()
layers = [
ResBlock(
ch + ich,
time_embed_dim,
dropout,
out_channels=model_channels * mult,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
)
]
ch = model_channels * mult
if ds in attention_resolutions:
if num_head_channels == -1:
dim_head = ch // num_heads
else:
num_heads = ch // num_head_channels
dim_head = num_head_channels
if legacy:
#num_heads = 1
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
layers.append(
AttentionBlock(
ch,
use_checkpoint=use_checkpoint,
num_heads=num_heads_upsample,
num_head_channels=dim_head,
use_new_attention_order=use_new_attention_order,
) if not use_spatial_transformer else SpatialTransformer(
ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
)
)
if level and i == num_res_blocks:
out_ch = ch
layers.append(
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=out_ch,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
up=True,
)
if resblock_updown
else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
)
ds //= 2
self.output_blocks.append(TimestepEmbedSequential(*layers))
self._feature_size += ch
self.out = nn.Sequential(
normalization(ch),
nn.SiLU(),
zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
)
if self.predict_codebook_ids:
self.id_predictor = nn.Sequential(
normalization(ch),
conv_nd(dims, model_channels, n_embed, 1),
#nn.LogSoftmax(dim=1) # change to cross_entropy and produce non-normalized logits
)
def convert_to_fp16(self):
"""
Convert the torso of the model to float16.
"""
self.input_blocks.apply(convert_module_to_f16)
self.middle_block.apply(convert_module_to_f16)
self.output_blocks.apply(convert_module_to_f16)
def convert_to_fp32(self):
"""
Convert the torso of the model to float32.
"""
self.input_blocks.apply(convert_module_to_f32)
self.middle_block.apply(convert_module_to_f32)
self.output_blocks.apply(convert_module_to_f32)
def forward(self, x, timesteps=None, context=None, y=None,**kwargs):
"""
Apply the model to an input batch.
:param x: an [N x C x ...] Tensor of inputs.
:param timesteps: a 1-D batch of timesteps.
:param context: conditioning plugged in via crossattn
:param y: an [N] Tensor of labels, if class-conditional.
:return: an [N x C x ...] Tensor of outputs.
"""
assert (y is not None) == (
self.num_classes is not None
), "must specify y if and only if the model is class-conditional"
hs = []
t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
emb = self.time_embed(t_emb)
if self.num_classes is not None:
assert y.shape == (x.shape[0],)
emb = emb + self.label_emb(y)
h = x.type(self.dtype)
for module in self.input_blocks:
h = module(h, emb, context)
hs.append(h)
h = self.middle_block(h, emb, context)
for module in self.output_blocks:
h = th.cat([h, hs.pop()], dim=1)
h = module(h, emb, context)
h = h.type(x.dtype)
if self.predict_codebook_ids:
return self.id_predictor(h)
else:
return self.out(h)
class EncoderUNetModel(nn.Module):
"""
The half UNet model with attention and timestep embedding.
For usage, see UNet.
"""
def __init__(
self,
image_size,
in_channels,
model_channels,
out_channels,
num_res_blocks,
attention_resolutions,
dropout=0,
channel_mult=(1, 2, 4, 8),
conv_resample=True,
dims=2,
use_checkpoint=False,
use_fp16=False,
num_heads=1,
num_head_channels=-1,
num_heads_upsample=-1,
use_scale_shift_norm=False,
resblock_updown=False,
use_new_attention_order=False,
pool="adaptive",
*args,
**kwargs
):
super().__init__()
if num_heads_upsample == -1:
num_heads_upsample = num_heads
self.in_channels = in_channels
self.model_channels = model_channels
self.out_channels = out_channels
self.num_res_blocks = num_res_blocks
self.attention_resolutions = attention_resolutions
self.dropout = dropout
self.channel_mult = channel_mult
self.conv_resample = conv_resample
self.use_checkpoint = use_checkpoint
self.dtype = th.float16 if use_fp16 else th.float32
self.num_heads = num_heads
self.num_head_channels = num_head_channels
self.num_heads_upsample = num_heads_upsample
time_embed_dim = model_channels * 4
self.time_embed = nn.Sequential(
linear(model_channels, time_embed_dim),
nn.SiLU(),
linear(time_embed_dim, time_embed_dim),
)
self.input_blocks = nn.ModuleList(
[
TimestepEmbedSequential(
conv_nd(dims, in_channels, model_channels, 3, padding=1)
)
]
)
self._feature_size = model_channels
input_block_chans = [model_channels]
ch = model_channels
ds = 1
for level, mult in enumerate(channel_mult):
for _ in range(num_res_blocks):
layers = [
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=mult * model_channels,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
)
]
ch = mult * model_channels
if ds in attention_resolutions:
layers.append(
AttentionBlock(
ch,
use_checkpoint=use_checkpoint,
num_heads=num_heads,
num_head_channels=num_head_channels,
use_new_attention_order=use_new_attention_order,
)
)
self.input_blocks.append(TimestepEmbedSequential(*layers))
self._feature_size += ch
input_block_chans.append(ch)
if level != len(channel_mult) - 1:
out_ch = ch
self.input_blocks.append(
TimestepEmbedSequential(
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=out_ch,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
down=True,
)
if resblock_updown
else Downsample(
ch, conv_resample, dims=dims, out_channels=out_ch
)
)
)
ch = out_ch
input_block_chans.append(ch)
ds *= 2
self._feature_size += ch
self.middle_block = TimestepEmbedSequential(
ResBlock(
ch,
time_embed_dim,
dropout,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
),
AttentionBlock(
ch,
use_checkpoint=use_checkpoint,
num_heads=num_heads,
num_head_channels=num_head_channels,
use_new_attention_order=use_new_attention_order,
),
ResBlock(
ch,
time_embed_dim,
dropout,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
),
)
self._feature_size += ch
self.pool = pool
if pool == "adaptive":
self.out = nn.Sequential(
normalization(ch),
nn.SiLU(),
nn.AdaptiveAvgPool2d((1, 1)),
zero_module(conv_nd(dims, ch, out_channels, 1)),
nn.Flatten(),
)
elif pool == "attention":
assert num_head_channels != -1
self.out = nn.Sequential(
normalization(ch),
nn.SiLU(),
AttentionPool2d(
(image_size // ds), ch, num_head_channels, out_channels
),
)
elif pool == "spatial":
self.out = nn.Sequential(
nn.Linear(self._feature_size, 2048),
nn.ReLU(),
nn.Linear(2048, self.out_channels),
)
elif pool == "spatial_v2":
self.out = nn.Sequential(
nn.Linear(self._feature_size, 2048),
normalization(2048),
nn.SiLU(),
nn.Linear(2048, self.out_channels),
)
else:
raise NotImplementedError(f"Unexpected {pool} pooling")
def convert_to_fp16(self):
"""
Convert the torso of the model to float16.
"""
self.input_blocks.apply(convert_module_to_f16)
self.middle_block.apply(convert_module_to_f16)
def convert_to_fp32(self):
"""
Convert the torso of the model to float32.
"""
self.input_blocks.apply(convert_module_to_f32)
self.middle_block.apply(convert_module_to_f32)
def forward(self, x, timesteps):
"""
Apply the model to an input batch.
:param x: an [N x C x ...] Tensor of inputs.
:param timesteps: a 1-D batch of timesteps.
:return: an [N x K] Tensor of outputs.
"""
emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
results = []
h = x.type(self.dtype)
for module in self.input_blocks:
h = module(h, emb)
if self.pool.startswith("spatial"):
results.append(h.type(x.dtype).mean(dim=(2, 3)))
h = self.middle_block(h, emb)
if self.pool.startswith("spatial"):
results.append(h.type(x.dtype).mean(dim=(2, 3)))
h = th.cat(results, axis=-1)
return self.out(h)
else:
h = h.type(x.dtype)
return self.out(h)

267
ldm/modules/diffusionmodules/util.py Normal file
View File

@ -0,0 +1,267 @@
# adopted from
# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
# and
# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
# and
# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
#
# thanks!
import os
import math
import torch
import torch.nn as nn
import numpy as np
from einops import repeat
from ldm.util import instantiate_from_config
def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
if schedule == "linear":
betas = (
torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
)
elif schedule == "cosine":
timesteps = (
torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
)
alphas = timesteps / (1 + cosine_s) * np.pi / 2
alphas = torch.cos(alphas).pow(2)
alphas = alphas / alphas[0]
betas = 1 - alphas[1:] / alphas[:-1]
betas = np.clip(betas, a_min=0, a_max=0.999)
elif schedule == "sqrt_linear":
betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
elif schedule == "sqrt":
betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
else:
raise ValueError(f"schedule '{schedule}' unknown.")
return betas.numpy()
def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True):
if ddim_discr_method == 'uniform':
c = num_ddpm_timesteps // num_ddim_timesteps
ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
elif ddim_discr_method == 'quad':
ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int)
else:
raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"')
# assert ddim_timesteps.shape[0] == num_ddim_timesteps
# add one to get the final alpha values right (the ones from first scale to data during sampling)
steps_out = ddim_timesteps + 1
if verbose:
print(f'Selected timesteps for ddim sampler: {steps_out}')
return steps_out
def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
# select alphas for computing the variance schedule
alphas = alphacums[ddim_timesteps]
alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
# according the the formula provided in https://arxiv.org/abs/2010.02502
sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev))
if verbose:
print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}')
print(f'For the chosen value of eta, which is {eta}, '
f'this results in the following sigma_t schedule for ddim sampler {sigmas}')
return sigmas, alphas, alphas_prev
def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
"""
Create a beta schedule that discretizes the given alpha_t_bar function,
which defines the cumulative product of (1-beta) over time from t = [0,1].
:param num_diffusion_timesteps: the number of betas to produce.
:param alpha_bar: a lambda that takes an argument t from 0 to 1 and
produces the cumulative product of (1-beta) up to that
part of the diffusion process.
:param max_beta: the maximum beta to use; use values lower than 1 to
prevent singularities.
"""
betas = []
for i in range(num_diffusion_timesteps):
t1 = i / num_diffusion_timesteps
t2 = (i + 1) / num_diffusion_timesteps
betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
return np.array(betas)
def extract_into_tensor(a, t, x_shape):
b, *_ = t.shape
out = a.gather(-1, t)
return out.reshape(b, *((1,) * (len(x_shape) - 1)))
def checkpoint(func, inputs, params, flag):
"""
Evaluate a function without caching intermediate activations, allowing for
reduced memory at the expense of extra compute in the backward pass.
:param func: the function to evaluate.
:param inputs: the argument sequence to pass to `func`.
:param params: a sequence of parameters `func` depends on but does not
explicitly take as arguments.
:param flag: if False, disable gradient checkpointing.
"""
if flag:
args = tuple(inputs) + tuple(params)
return CheckpointFunction.apply(func, len(inputs), *args)
else:
return func(*inputs)
class CheckpointFunction(torch.autograd.Function):
@staticmethod
def forward(ctx, run_function, length, *args):
ctx.run_function = run_function
ctx.input_tensors = list(args[:length])
ctx.input_params = list(args[length:])
with torch.no_grad():
output_tensors = ctx.run_function(*ctx.input_tensors)
return output_tensors
@staticmethod
def backward(ctx, *output_grads):
ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
with torch.enable_grad():
# Fixes a bug where the first op in run_function modifies the
# Tensor storage in place, which is not allowed for detach()'d
# Tensors.
shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
output_tensors = ctx.run_function(*shallow_copies)
input_grads = torch.autograd.grad(
output_tensors,
ctx.input_tensors + ctx.input_params,
output_grads,
allow_unused=True,
)
del ctx.input_tensors
del ctx.input_params
del output_tensors
return (None, None) + input_grads
def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
"""
Create sinusoidal timestep embeddings.
:param timesteps: a 1-D Tensor of N indices, one per batch element.
These may be fractional.
:param dim: the dimension of the output.
:param max_period: controls the minimum frequency of the embeddings.
:return: an [N x dim] Tensor of positional embeddings.
"""
if not repeat_only:
half = dim // 2
freqs = torch.exp(
-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
).to(device=timesteps.device)
args = timesteps[:, None].float() * freqs[None]
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
if dim % 2:
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
else:
embedding = repeat(timesteps, 'b -> b d', d=dim)
return embedding
def zero_module(module):
"""
Zero out the parameters of a module and return it.
"""
for p in module.parameters():
p.detach().zero_()
return module
def scale_module(module, scale):
"""
Scale the parameters of a module and return it.
"""
for p in module.parameters():
p.detach().mul_(scale)
return module
def mean_flat(tensor):
"""
Take the mean over all non-batch dimensions.
"""
return tensor.mean(dim=list(range(1, len(tensor.shape))))
def normalization(channels):
"""
Make a standard normalization layer.
:param channels: number of input channels.
:return: an nn.Module for normalization.
"""
return GroupNorm32(32, channels)
# PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
class SiLU(nn.Module):
def forward(self, x):
return x * torch.sigmoid(x)
class GroupNorm32(nn.GroupNorm):
def forward(self, x):
return super().forward(x.float()).type(x.dtype)
def conv_nd(dims, *args, **kwargs):
"""
Create a 1D, 2D, or 3D convolution module.
"""
if dims == 1:
return nn.Conv1d(*args, **kwargs)
elif dims == 2:
return nn.Conv2d(*args, **kwargs)
elif dims == 3:
return nn.Conv3d(*args, **kwargs)
raise ValueError(f"unsupported dimensions: {dims}")
def linear(*args, **kwargs):
"""
Create a linear module.
"""
return nn.Linear(*args, **kwargs)
def avg_pool_nd(dims, *args, **kwargs):
"""
Create a 1D, 2D, or 3D average pooling module.
"""
if dims == 1:
return nn.AvgPool1d(*args, **kwargs)
elif dims == 2:
return nn.AvgPool2d(*args, **kwargs)
elif dims == 3:
return nn.AvgPool3d(*args, **kwargs)
raise ValueError(f"unsupported dimensions: {dims}")
class HybridConditioner(nn.Module):
def __init__(self, c_concat_config, c_crossattn_config):
super().__init__()
self.concat_conditioner = instantiate_from_config(c_concat_config)
self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
def forward(self, c_concat, c_crossattn):
c_concat = self.concat_conditioner(c_concat)
c_crossattn = self.crossattn_conditioner(c_crossattn)
return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]}
def noise_like(shape, device, repeat=False):
repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
noise = lambda: torch.randn(shape, device=device)
return repeat_noise() if repeat else noise()

0
ldm/modules/distributions/__init__.py Normal file
View File

92
ldm/modules/distributions/distributions.py Normal file
View File

@ -0,0 +1,92 @@
import torch
import numpy as np
class AbstractDistribution:
def sample(self):
raise NotImplementedError()
def mode(self):
raise NotImplementedError()
class DiracDistribution(AbstractDistribution):
def __init__(self, value):
self.value = value
def sample(self):
return self.value
def mode(self):
return self.value
class DiagonalGaussianDistribution(object):
def __init__(self, parameters, deterministic=False):
self.parameters = parameters
self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
self.deterministic = deterministic
self.std = torch.exp(0.5 * self.logvar)
self.var = torch.exp(self.logvar)
if self.deterministic:
self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
def sample(self):
x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device)
return x
def kl(self, other=None):
if self.deterministic:
return torch.Tensor([0.])
else:
if other is None:
return 0.5 * torch.sum(torch.pow(self.mean, 2)
+ self.var - 1.0 - self.logvar,
dim=[1, 2, 3])
else:
return 0.5 * torch.sum(
torch.pow(self.mean - other.mean, 2) / other.var
+ self.var / other.var - 1.0 - self.logvar + other.logvar,
dim=[1, 2, 3])
def nll(self, sample, dims=[1,2,3]):
if self.deterministic:
return torch.Tensor([0.])
logtwopi = np.log(2.0 * np.pi)
return 0.5 * torch.sum(
logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
dim=dims)
def mode(self):
return self.mean
def normal_kl(mean1, logvar1, mean2, logvar2):
"""
source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
Compute the KL divergence between two gaussians.
Shapes are automatically broadcasted, so batches can be compared to
scalars, among other use cases.
"""
tensor = None
for obj in (mean1, logvar1, mean2, logvar2):
if isinstance(obj, torch.Tensor):
tensor = obj
break
assert tensor is not None, "at least one argument must be a Tensor"
# Force variances to be Tensors. Broadcasting helps convert scalars to
# Tensors, but it does not work for torch.exp().
logvar1, logvar2 = [
x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
for x in (logvar1, logvar2)
]
return 0.5 * (
-1.0
+ logvar2
- logvar1
+ torch.exp(logvar1 - logvar2)
+ ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
)

76
ldm/modules/ema.py Normal file
View File

@ -0,0 +1,76 @@
import torch
from torch import nn
class LitEma(nn.Module):
def __init__(self, model, decay=0.9999, use_num_upates=True):
super().__init__()
if decay < 0.0 or decay > 1.0:
raise ValueError('Decay must be between 0 and 1')
self.m_name2s_name = {}
self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))
self.register_buffer('num_updates', torch.tensor(0,dtype=torch.int) if use_num_upates
else torch.tensor(-1,dtype=torch.int))
for name, p in model.named_parameters():
if p.requires_grad:
#remove as '.'-character is not allowed in buffers
s_name = name.replace('.','')
self.m_name2s_name.update({name:s_name})
self.register_buffer(s_name,p.clone().detach().data)
self.collected_params = []
def forward(self,model):
decay = self.decay
if self.num_updates >= 0:
self.num_updates += 1
decay = min(self.decay,(1 + self.num_updates) / (10 + self.num_updates))
one_minus_decay = 1.0 - decay
with torch.no_grad():
m_param = dict(model.named_parameters())
shadow_params = dict(self.named_buffers())
for key in m_param:
if m_param[key].requires_grad:
sname = self.m_name2s_name[key]
shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
else:
assert not key in self.m_name2s_name
def copy_to(self, model):
m_param = dict(model.named_parameters())
shadow_params = dict(self.named_buffers())
for key in m_param:
if m_param[key].requires_grad:
m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
else:
assert not key in self.m_name2s_name
def store(self, parameters):
"""
Save the current parameters for restoring later.
Args:
parameters: Iterable of `torch.nn.Parameter`; the parameters to be
temporarily stored.
"""
self.collected_params = [param.clone() for param in parameters]
def restore(self, parameters):
"""
Restore the parameters stored with the `store` method.
Useful to validate the model with EMA parameters without affecting the
original optimization process. Store the parameters before the
`copy_to` method. After validation (or model saving), use this to
restore the former parameters.
Args:
parameters: Iterable of `torch.nn.Parameter`; the parameters to be
updated with the stored parameters.
"""
for c_param, param in zip(self.collected_params, parameters):
param.data.copy_(c_param.data)

0
ldm/modules/encoders/__init__.py Normal file
View File

234
ldm/modules/encoders/modules.py Normal file
View File

@ -0,0 +1,234 @@
import torch
import torch.nn as nn
from functools import partial
import clip
from einops import rearrange, repeat
from transformers import CLIPTokenizer, CLIPTextModel
import kornia
from ldm.modules.x_transformer import Encoder, TransformerWrapper # TODO: can we directly rely on lucidrains code and simply add this as a reuirement? --> test
class AbstractEncoder(nn.Module):
def __init__(self):
super().__init__()
def encode(self, *args, **kwargs):
raise NotImplementedError
class ClassEmbedder(nn.Module):
def __init__(self, embed_dim, n_classes=1000, key='class'):
super().__init__()
self.key = key
self.embedding = nn.Embedding(n_classes, embed_dim)
def forward(self, batch, key=None):
if key is None:
key = self.key
# this is for use in crossattn
c = batch[key][:, None]
c = self.embedding(c)
return c
class TransformerEmbedder(AbstractEncoder):
"""Some transformer encoder layers"""
def __init__(self, n_embed, n_layer, vocab_size, max_seq_len=77, device="cuda"):
super().__init__()
self.device = device
self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len,
attn_layers=Encoder(dim=n_embed, depth=n_layer))
def forward(self, tokens):
tokens = tokens.to(self.device) # meh
z = self.transformer(tokens, return_embeddings=True)
return z
def encode(self, x):
return self(x)
class BERTTokenizer(AbstractEncoder):
""" Uses a pretrained BERT tokenizer by huggingface. Vocab size: 30522 (?)"""
def __init__(self, device="cuda", vq_interface=True, max_length=77):
super().__init__()
from transformers import BertTokenizerFast # TODO: add to reuquirements
self.tokenizer = BertTokenizerFast.from_pretrained("bert-base-uncased")
self.device = device
self.vq_interface = vq_interface
self.max_length = max_length
def forward(self, text):
batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
tokens = batch_encoding["input_ids"].to(self.device)
return tokens
@torch.no_grad()
def encode(self, text):
tokens = self(text)
if not self.vq_interface:
return tokens
return None, None, [None, None, tokens]
def decode(self, text):
return text
class BERTEmbedder(AbstractEncoder):
"""Uses the BERT tokenizr model and add some transformer encoder layers"""
def __init__(self, n_embed, n_layer, vocab_size=30522, max_seq_len=77,
device="cuda",use_tokenizer=True, embedding_dropout=0.0):
super().__init__()
self.use_tknz_fn = use_tokenizer
if self.use_tknz_fn:
self.tknz_fn = BERTTokenizer(vq_interface=False, max_length=max_seq_len)
self.device = device
self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len,
attn_layers=Encoder(dim=n_embed, depth=n_layer),
emb_dropout=embedding_dropout)
def forward(self, text):
if self.use_tknz_fn:
tokens = self.tknz_fn(text)#.to(self.device)
else:
tokens = text
z = self.transformer(tokens, return_embeddings=True)
return z
def encode(self, text):
# output of length 77
return self(text)
class SpatialRescaler(nn.Module):
def __init__(self,
n_stages=1,
method='bilinear',
multiplier=0.5,
in_channels=3,
out_channels=None,
bias=False):
super().__init__()
self.n_stages = n_stages
assert self.n_stages >= 0
assert method in ['nearest','linear','bilinear','trilinear','bicubic','area']
self.multiplier = multiplier
self.interpolator = partial(torch.nn.functional.interpolate, mode=method)
self.remap_output = out_channels is not None
if self.remap_output:
print(f'Spatial Rescaler mapping from {in_channels} to {out_channels} channels after resizing.')
self.channel_mapper = nn.Conv2d(in_channels,out_channels,1,bias=bias)
def forward(self,x):
for stage in range(self.n_stages):
x = self.interpolator(x, scale_factor=self.multiplier)
if self.remap_output:
x = self.channel_mapper(x)
return x
def encode(self, x):
return self(x)
class FrozenCLIPEmbedder(AbstractEncoder):
"""Uses the CLIP transformer encoder for text (from Hugging Face)"""
def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77):
super().__init__()
self.tokenizer = CLIPTokenizer.from_pretrained(version)
self.transformer = CLIPTextModel.from_pretrained(version)
self.device = device
self.max_length = max_length
self.freeze()
def freeze(self):
self.transformer = self.transformer.eval()
for param in self.parameters():
param.requires_grad = False
def forward(self, text):
batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
tokens = batch_encoding["input_ids"].to(self.device)
outputs = self.transformer(input_ids=tokens)
z = outputs.last_hidden_state
return z
def encode(self, text):
return self(text)
class FrozenCLIPTextEmbedder(nn.Module):
"""
Uses the CLIP transformer encoder for text.
"""
def __init__(self, version='ViT-L/14', device="cuda", max_length=77, n_repeat=1, normalize=True):
super().__init__()
self.model, _ = clip.load(version, jit=False, device="cpu")
self.device = device
self.max_length = max_length
self.n_repeat = n_repeat
self.normalize = normalize
def freeze(self):
self.model = self.model.eval()
for param in self.parameters():
param.requires_grad = False
def forward(self, text):
tokens = clip.tokenize(text).to(self.device)
z = self.model.encode_text(tokens)
if self.normalize:
z = z / torch.linalg.norm(z, dim=1, keepdim=True)
return z
def encode(self, text):
z = self(text)
if z.ndim==2:
z = z[:, None, :]
z = repeat(z, 'b 1 d -> b k d', k=self.n_repeat)
return z
class FrozenClipImageEmbedder(nn.Module):
"""
Uses the CLIP image encoder.
"""
def __init__(
self,
model,
jit=False,
device='cuda' if torch.cuda.is_available() else 'cpu',
antialias=False,
):
super().__init__()
self.model, _ = clip.load(name=model, device=device, jit=jit)
self.antialias = antialias
self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
def preprocess(self, x):
# normalize to [0,1]
x = kornia.geometry.resize(x, (224, 224),
interpolation='bicubic',align_corners=True,
antialias=self.antialias)
x = (x + 1.) / 2.
# renormalize according to clip
x = kornia.enhance.normalize(x, self.mean, self.std)
return x
def forward(self, x):
# x is assumed to be in range [-1,1]
return self.model.encode_image(self.preprocess(x))
if __name__ == "__main__":
from ldm.util import count_params
model = FrozenCLIPEmbedder()
count_params(model, verbose=True)

2
ldm/modules/image_degradation/__init__.py Normal file
View File

@ -0,0 +1,2 @@
from ldm.modules.image_degradation.bsrgan import degradation_bsrgan_variant as degradation_fn_bsr
from ldm.modules.image_degradation.bsrgan_light import degradation_bsrgan_variant as degradation_fn_bsr_light

730
ldm/modules/image_degradation/bsrgan.py Normal file
View File

@ -0,0 +1,730 @@
# -*- coding: utf-8 -*-
"""
# --------------------------------------------
# Super-Resolution
# --------------------------------------------
#
# Kai Zhang (cskaizhang@gmail.com)
# https://github.com/cszn
# From 2019/03--2021/08
# --------------------------------------------
"""
import numpy as np
import cv2
import torch
from functools import partial
import random
from scipy import ndimage
import scipy
import scipy.stats as ss
from scipy.interpolate import interp2d
from scipy.linalg import orth
import albumentations
import ldm.modules.image_degradation.utils_image as util
def modcrop_np(img, sf):
'''
Args:
img: numpy image, WxH or WxHxC
sf: scale factor
Return:
cropped image
'''
w, h = img.shape[:2]
im = np.copy(img)
return im[:w - w % sf, :h - h % sf, ...]
"""
# --------------------------------------------
# anisotropic Gaussian kernels
# --------------------------------------------
"""
def analytic_kernel(k):
"""Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)"""
k_size = k.shape[0]
# Calculate the big kernels size
big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2))
# Loop over the small kernel to fill the big one
for r in range(k_size):
for c in range(k_size):
big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k
# Crop the edges of the big kernel to ignore very small values and increase run time of SR
crop = k_size // 2
cropped_big_k = big_k[crop:-crop, crop:-crop]
# Normalize to 1
return cropped_big_k / cropped_big_k.sum()
def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6):
""" generate an anisotropic Gaussian kernel
Args:
ksize : e.g., 15, kernel size
theta : [0, pi], rotation angle range
l1 : [0.1,50], scaling of eigenvalues
l2 : [0.1,l1], scaling of eigenvalues
If l1 = l2, will get an isotropic Gaussian kernel.
Returns:
k : kernel
"""
v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.]))
V = np.array([[v[0], v[1]], [v[1], -v[0]]])
D = np.array([[l1, 0], [0, l2]])
Sigma = np.dot(np.dot(V, D), np.linalg.inv(V))
k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize)
return k
def gm_blur_kernel(mean, cov, size=15):
center = size / 2.0 + 0.5
k = np.zeros([size, size])
for y in range(size):
for x in range(size):
cy = y - center + 1
cx = x - center + 1
k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov)
k = k / np.sum(k)
return k
def shift_pixel(x, sf, upper_left=True):
"""shift pixel for super-resolution with different scale factors
Args:
x: WxHxC or WxH
sf: scale factor
upper_left: shift direction
"""
h, w = x.shape[:2]
shift = (sf - 1) * 0.5
xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0)
if upper_left:
x1 = xv + shift
y1 = yv + shift
else:
x1 = xv - shift
y1 = yv - shift
x1 = np.clip(x1, 0, w - 1)
y1 = np.clip(y1, 0, h - 1)
if x.ndim == 2:
x = interp2d(xv, yv, x)(x1, y1)
if x.ndim == 3:
for i in range(x.shape[-1]):
x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1)
return x
def blur(x, k):
'''
x: image, NxcxHxW
k: kernel, Nx1xhxw
'''
n, c = x.shape[:2]
p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2
x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate')
k = k.repeat(1, c, 1, 1)
k = k.view(-1, 1, k.shape[2], k.shape[3])
x = x.view(1, -1, x.shape[2], x.shape[3])
x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c)
x = x.view(n, c, x.shape[2], x.shape[3])
return x
def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0):
""""
# modified version of https://github.com/assafshocher/BlindSR_dataset_generator
# Kai Zhang
# min_var = 0.175 * sf # variance of the gaussian kernel will be sampled between min_var and max_var
# max_var = 2.5 * sf
"""
# Set random eigen-vals (lambdas) and angle (theta) for COV matrix
lambda_1 = min_var + np.random.rand() * (max_var - min_var)
lambda_2 = min_var + np.random.rand() * (max_var - min_var)
theta = np.random.rand() * np.pi # random theta
noise = -noise_level + np.random.rand(*k_size) * noise_level * 2
# Set COV matrix using Lambdas and Theta
LAMBDA = np.diag([lambda_1, lambda_2])
Q = np.array([[np.cos(theta), -np.sin(theta)],
[np.sin(theta), np.cos(theta)]])
SIGMA = Q @ LAMBDA @ Q.T
INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :]
# Set expectation position (shifting kernel for aligned image)
MU = k_size // 2 - 0.5 * (scale_factor - 1) # - 0.5 * (scale_factor - k_size % 2)
MU = MU[None, None, :, None]
# Create meshgrid for Gaussian
[X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1]))
Z = np.stack([X, Y], 2)[:, :, :, None]
# Calcualte Gaussian for every pixel of the kernel
ZZ = Z - MU
ZZ_t = ZZ.transpose(0, 1, 3, 2)
raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise)
# shift the kernel so it will be centered
# raw_kernel_centered = kernel_shift(raw_kernel, scale_factor)
# Normalize the kernel and return
# kernel = raw_kernel_centered / np.sum(raw_kernel_centered)
kernel = raw_kernel / np.sum(raw_kernel)
return kernel
def fspecial_gaussian(hsize, sigma):
hsize = [hsize, hsize]
siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0]
std = sigma
[x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1))
arg = -(x * x + y * y) / (2 * std * std)
h = np.exp(arg)
h[h < scipy.finfo(float).eps * h.max()] = 0
sumh = h.sum()
if sumh != 0:
h = h / sumh
return h
def fspecial_laplacian(alpha):
alpha = max([0, min([alpha, 1])])
h1 = alpha / (alpha + 1)
h2 = (1 - alpha) / (alpha + 1)
h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]]
h = np.array(h)
return h
def fspecial(filter_type, *args, **kwargs):
'''
python code from:
https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py
'''
if filter_type == 'gaussian':
return fspecial_gaussian(*args, **kwargs)
if filter_type == 'laplacian':
return fspecial_laplacian(*args, **kwargs)
"""
# --------------------------------------------
# degradation models
# --------------------------------------------
"""
def bicubic_degradation(x, sf=3):
'''
Args:
x: HxWxC image, [0, 1]
sf: down-scale factor
Return:
bicubicly downsampled LR image
'''
x = util.imresize_np(x, scale=1 / sf)
return x
def srmd_degradation(x, k, sf=3):
''' blur + bicubic downsampling
Args:
x: HxWxC image, [0, 1]
k: hxw, double
sf: down-scale factor
Return:
downsampled LR image
Reference:
@inproceedings{zhang2018learning,
title={Learning a single convolutional super-resolution network for multiple degradations},
author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
pages={3262--3271},
year={2018}
}
'''
x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap') # 'nearest' | 'mirror'
x = bicubic_degradation(x, sf=sf)
return x
def dpsr_degradation(x, k, sf=3):
''' bicubic downsampling + blur
Args:
x: HxWxC image, [0, 1]
k: hxw, double
sf: down-scale factor
Return:
downsampled LR image
Reference:
@inproceedings{zhang2019deep,
title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels},
author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
pages={1671--1681},
year={2019}
}
'''
x = bicubic_degradation(x, sf=sf)
x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
return x
def classical_degradation(x, k, sf=3):
''' blur + downsampling
Args:
x: HxWxC image, [0, 1]/[0, 255]
k: hxw, double
sf: down-scale factor
Return:
downsampled LR image
'''
x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
# x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2))
st = 0
return x[st::sf, st::sf, ...]
def add_sharpening(img, weight=0.5, radius=50, threshold=10):
"""USM sharpening. borrowed from real-ESRGAN
Input image: I; Blurry image: B.
1. K = I + weight * (I - B)
2. Mask = 1 if abs(I - B) > threshold, else: 0
3. Blur mask:
4. Out = Mask * K + (1 - Mask) * I
Args:
img (Numpy array): Input image, HWC, BGR; float32, [0, 1].
weight (float): Sharp weight. Default: 1.
radius (float): Kernel size of Gaussian blur. Default: 50.
threshold (int):
"""
if radius % 2 == 0:
radius += 1
blur = cv2.GaussianBlur(img, (radius, radius), 0)
residual = img - blur
mask = np.abs(residual) * 255 > threshold
mask = mask.astype('float32')
soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0)
K = img + weight * residual
K = np.clip(K, 0, 1)
return soft_mask * K + (1 - soft_mask) * img
def add_blur(img, sf=4):
wd2 = 4.0 + sf
wd = 2.0 + 0.2 * sf
if random.random() < 0.5:
l1 = wd2 * random.random()
l2 = wd2 * random.random()
k = anisotropic_Gaussian(ksize=2 * random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2)
else:
k = fspecial('gaussian', 2 * random.randint(2, 11) + 3, wd * random.random())
img = ndimage.filters.convolve(img, np.expand_dims(k, axis=2), mode='mirror')
return img
def add_resize(img, sf=4):
rnum = np.random.rand()
if rnum > 0.8: # up
sf1 = random.uniform(1, 2)
elif rnum < 0.7: # down
sf1 = random.uniform(0.5 / sf, 1)
else:
sf1 = 1.0
img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3]))
img = np.clip(img, 0.0, 1.0)
return img
# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
# noise_level = random.randint(noise_level1, noise_level2)
# rnum = np.random.rand()
# if rnum > 0.6: # add color Gaussian noise
# img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
# elif rnum < 0.4: # add grayscale Gaussian noise
# img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
# else: # add noise
# L = noise_level2 / 255.
# D = np.diag(np.random.rand(3))
# U = orth(np.random.rand(3, 3))
# conv = np.dot(np.dot(np.transpose(U), D), U)
# img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
# img = np.clip(img, 0.0, 1.0)
# return img
def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
noise_level = random.randint(noise_level1, noise_level2)
rnum = np.random.rand()
if rnum > 0.6: # add color Gaussian noise
img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
elif rnum < 0.4: # add grayscale Gaussian noise
img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
else: # add noise
L = noise_level2 / 255.
D = np.diag(np.random.rand(3))
U = orth(np.random.rand(3, 3))
conv = np.dot(np.dot(np.transpose(U), D), U)
img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
img = np.clip(img, 0.0, 1.0)
return img
def add_speckle_noise(img, noise_level1=2, noise_level2=25):
noise_level = random.randint(noise_level1, noise_level2)
img = np.clip(img, 0.0, 1.0)
rnum = random.random()
if rnum > 0.6:
img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
elif rnum < 0.4:
img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
else:
L = noise_level2 / 255.
D = np.diag(np.random.rand(3))
U = orth(np.random.rand(3, 3))
conv = np.dot(np.dot(np.transpose(U), D), U)
img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
img = np.clip(img, 0.0, 1.0)
return img
def add_Poisson_noise(img):
img = np.clip((img * 255.0).round(), 0, 255) / 255.
vals = 10 ** (2 * random.random() + 2.0) # [2, 4]
if random.random() < 0.5:
img = np.random.poisson(img * vals).astype(np.float32) / vals
else:
img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114])
img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255.
noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray
img += noise_gray[:, :, np.newaxis]
img = np.clip(img, 0.0, 1.0)
return img
def add_JPEG_noise(img):
quality_factor = random.randint(30, 95)
img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR)
result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor])
img = cv2.imdecode(encimg, 1)
img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB)
return img
def random_crop(lq, hq, sf=4, lq_patchsize=64):
h, w = lq.shape[:2]
rnd_h = random.randint(0, h - lq_patchsize)
rnd_w = random.randint(0, w - lq_patchsize)
lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :]
rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf)
hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :]
return lq, hq
def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
"""
This is the degradation model of BSRGAN from the paper
"Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
----------
img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
sf: scale factor
isp_model: camera ISP model
Returns
-------
img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
"""
isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
sf_ori = sf
h1, w1 = img.shape[:2]
img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop
h, w = img.shape[:2]
if h < lq_patchsize * sf or w < lq_patchsize * sf:
raise ValueError(f'img size ({h1}X{w1}) is too small!')
hq = img.copy()
if sf == 4 and random.random() < scale2_prob: # downsample1
if np.random.rand() < 0.5:
img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])),
interpolation=random.choice([1, 2, 3]))
else:
img = util.imresize_np(img, 1 / 2, True)
img = np.clip(img, 0.0, 1.0)
sf = 2
shuffle_order = random.sample(range(7), 7)
idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
if idx1 > idx2: # keep downsample3 last
shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
for i in shuffle_order:
if i == 0:
img = add_blur(img, sf=sf)
elif i == 1:
img = add_blur(img, sf=sf)
elif i == 2:
a, b = img.shape[1], img.shape[0]
# downsample2
if random.random() < 0.75:
sf1 = random.uniform(1, 2 * sf)
img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])),
interpolation=random.choice([1, 2, 3]))
else:
k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
k_shifted = shift_pixel(k, sf)
k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel
img = ndimage.filters.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror')
img = img[0::sf, 0::sf, ...] # nearest downsampling
img = np.clip(img, 0.0, 1.0)
elif i == 3:
# downsample3
img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
img = np.clip(img, 0.0, 1.0)
elif i == 4:
# add Gaussian noise
img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
elif i == 5:
# add JPEG noise
if random.random() < jpeg_prob:
img = add_JPEG_noise(img)
elif i == 6:
# add processed camera sensor noise
if random.random() < isp_prob and isp_model is not None:
with torch.no_grad():
img, hq = isp_model.forward(img.copy(), hq)
# add final JPEG compression noise
img = add_JPEG_noise(img)
# random crop
img, hq = random_crop(img, hq, sf_ori, lq_patchsize)
return img, hq
# todo no isp_model?
def degradation_bsrgan_variant(image, sf=4, isp_model=None):
"""
This is the degradation model of BSRGAN from the paper
"Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
----------
sf: scale factor
isp_model: camera ISP model
Returns
-------
img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
"""
image = util.uint2single(image)
isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
sf_ori = sf
h1, w1 = image.shape[:2]
image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop
h, w = image.shape[:2]
hq = image.copy()
if sf == 4 and random.random() < scale2_prob: # downsample1
if np.random.rand() < 0.5:
image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])),
interpolation=random.choice([1, 2, 3]))
else:
image = util.imresize_np(image, 1 / 2, True)
image = np.clip(image, 0.0, 1.0)
sf = 2
shuffle_order = random.sample(range(7), 7)
idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
if idx1 > idx2: # keep downsample3 last
shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
for i in shuffle_order:
if i == 0:
image = add_blur(image, sf=sf)
elif i == 1:
image = add_blur(image, sf=sf)
elif i == 2:
a, b = image.shape[1], image.shape[0]
# downsample2
if random.random() < 0.75:
sf1 = random.uniform(1, 2 * sf)
image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])),
interpolation=random.choice([1, 2, 3]))
else:
k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
k_shifted = shift_pixel(k, sf)
k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel
image = ndimage.filters.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror')
image = image[0::sf, 0::sf, ...] # nearest downsampling
image = np.clip(image, 0.0, 1.0)
elif i == 3:
# downsample3
image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
image = np.clip(image, 0.0, 1.0)
elif i == 4:
# add Gaussian noise
image = add_Gaussian_noise(image, noise_level1=2, noise_level2=25)
elif i == 5:
# add JPEG noise
if random.random() < jpeg_prob:
image = add_JPEG_noise(image)
# elif i == 6:
# # add processed camera sensor noise
# if random.random() < isp_prob and isp_model is not None:
# with torch.no_grad():
# img, hq = isp_model.forward(img.copy(), hq)
# add final JPEG compression noise
image = add_JPEG_noise(image)
image = util.single2uint(image)
example = {"image":image}
return example
# TODO incase there is a pickle error one needs to replace a += x with a = a + x in add_speckle_noise etc...
def degradation_bsrgan_plus(img, sf=4, shuffle_prob=0.5, use_sharp=True, lq_patchsize=64, isp_model=None):
"""
This is an extended degradation model by combining
the degradation models of BSRGAN and Real-ESRGAN
----------
img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
sf: scale factor
use_shuffle: the degradation shuffle
use_sharp: sharpening the img
Returns
-------
img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
"""
h1, w1 = img.shape[:2]
img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop
h, w = img.shape[:2]
if h < lq_patchsize * sf or w < lq_patchsize * sf:
raise ValueError(f'img size ({h1}X{w1}) is too small!')
if use_sharp:
img = add_sharpening(img)
hq = img.copy()
if random.random() < shuffle_prob:
shuffle_order = random.sample(range(13), 13)
else:
shuffle_order = list(range(13))
# local shuffle for noise, JPEG is always the last one
shuffle_order[2:6] = random.sample(shuffle_order[2:6], len(range(2, 6)))
shuffle_order[9:13] = random.sample(shuffle_order[9:13], len(range(9, 13)))
poisson_prob, speckle_prob, isp_prob = 0.1, 0.1, 0.1
for i in shuffle_order:
if i == 0:
img = add_blur(img, sf=sf)
elif i == 1:
img = add_resize(img, sf=sf)
elif i == 2:
img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
elif i == 3:
if random.random() < poisson_prob:
img = add_Poisson_noise(img)
elif i == 4:
if random.random() < speckle_prob:
img = add_speckle_noise(img)
elif i == 5:
if random.random() < isp_prob and isp_model is not None:
with torch.no_grad():
img, hq = isp_model.forward(img.copy(), hq)
elif i == 6:
img = add_JPEG_noise(img)
elif i == 7:
img = add_blur(img, sf=sf)
elif i == 8:
img = add_resize(img, sf=sf)
elif i == 9:
img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
elif i == 10:
if random.random() < poisson_prob:
img = add_Poisson_noise(img)
elif i == 11:
if random.random() < speckle_prob:
img = add_speckle_noise(img)
elif i == 12:
if random.random() < isp_prob and isp_model is not None:
with torch.no_grad():
img, hq = isp_model.forward(img.copy(), hq)
else:
print('check the shuffle!')
# resize to desired size
img = cv2.resize(img, (int(1 / sf * hq.shape[1]), int(1 / sf * hq.shape[0])),
interpolation=random.choice([1, 2, 3]))
# add final JPEG compression noise
img = add_JPEG_noise(img)
# random crop
img, hq = random_crop(img, hq, sf, lq_patchsize)
return img, hq
if __name__ == '__main__':
print("hey")
img = util.imread_uint('utils/test.png', 3)
print(img)
img = util.uint2single(img)
print(img)
img = img[:448, :448]
h = img.shape[0] // 4
print("resizing to", h)
sf = 4
deg_fn = partial(degradation_bsrgan_variant, sf=sf)
for i in range(20):
print(i)
img_lq = deg_fn(img)
print(img_lq)
img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img)["image"]
print(img_lq.shape)
print("bicubic", img_lq_bicubic.shape)
print(img_hq.shape)
lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
interpolation=0)
lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
interpolation=0)
img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1)
util.imsave(img_concat, str(i) + '.png')

650
ldm/modules/image_degradation/bsrgan_light.py Normal file
View File

@ -0,0 +1,650 @@
# -*- coding: utf-8 -*-
import numpy as np
import cv2
import torch
from functools import partial
import random
from scipy import ndimage
import scipy
import scipy.stats as ss
from scipy.interpolate import interp2d
from scipy.linalg import orth
import albumentations
import ldm.modules.image_degradation.utils_image as util
"""
# --------------------------------------------
# Super-Resolution
# --------------------------------------------
#
# Kai Zhang (cskaizhang@gmail.com)
# https://github.com/cszn
# From 2019/03--2021/08
# --------------------------------------------
"""
def modcrop_np(img, sf):
'''
Args:
img: numpy image, WxH or WxHxC
sf: scale factor
Return:
cropped image
'''
w, h = img.shape[:2]
im = np.copy(img)
return im[:w - w % sf, :h - h % sf, ...]
"""
# --------------------------------------------
# anisotropic Gaussian kernels
# --------------------------------------------
"""
def analytic_kernel(k):
"""Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)"""
k_size = k.shape[0]
# Calculate the big kernels size
big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2))
# Loop over the small kernel to fill the big one
for r in range(k_size):
for c in range(k_size):
big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k
# Crop the edges of the big kernel to ignore very small values and increase run time of SR
crop = k_size // 2
cropped_big_k = big_k[crop:-crop, crop:-crop]
# Normalize to 1
return cropped_big_k / cropped_big_k.sum()
def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6):
""" generate an anisotropic Gaussian kernel
Args:
ksize : e.g., 15, kernel size
theta : [0, pi], rotation angle range
l1 : [0.1,50], scaling of eigenvalues
l2 : [0.1,l1], scaling of eigenvalues
If l1 = l2, will get an isotropic Gaussian kernel.
Returns:
k : kernel
"""
v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.]))
V = np.array([[v[0], v[1]], [v[1], -v[0]]])
D = np.array([[l1, 0], [0, l2]])
Sigma = np.dot(np.dot(V, D), np.linalg.inv(V))
k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize)
return k
def gm_blur_kernel(mean, cov, size=15):
center = size / 2.0 + 0.5
k = np.zeros([size, size])
for y in range(size):
for x in range(size):
cy = y - center + 1
cx = x - center + 1
k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov)
k = k / np.sum(k)
return k
def shift_pixel(x, sf, upper_left=True):
"""shift pixel for super-resolution with different scale factors
Args:
x: WxHxC or WxH
sf: scale factor
upper_left: shift direction
"""
h, w = x.shape[:2]
shift = (sf - 1) * 0.5
xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0)
if upper_left:
x1 = xv + shift
y1 = yv + shift
else:
x1 = xv - shift
y1 = yv - shift
x1 = np.clip(x1, 0, w - 1)
y1 = np.clip(y1, 0, h - 1)
if x.ndim == 2:
x = interp2d(xv, yv, x)(x1, y1)
if x.ndim == 3:
for i in range(x.shape[-1]):
x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1)
return x
def blur(x, k):
'''
x: image, NxcxHxW
k: kernel, Nx1xhxw
'''
n, c = x.shape[:2]
p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2
x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate')
k = k.repeat(1, c, 1, 1)
k = k.view(-1, 1, k.shape[2], k.shape[3])
x = x.view(1, -1, x.shape[2], x.shape[3])
x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c)
x = x.view(n, c, x.shape[2], x.shape[3])
return x
def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0):
""""
# modified version of https://github.com/assafshocher/BlindSR_dataset_generator
# Kai Zhang
# min_var = 0.175 * sf # variance of the gaussian kernel will be sampled between min_var and max_var
# max_var = 2.5 * sf
"""
# Set random eigen-vals (lambdas) and angle (theta) for COV matrix
lambda_1 = min_var + np.random.rand() * (max_var - min_var)
lambda_2 = min_var + np.random.rand() * (max_var - min_var)
theta = np.random.rand() * np.pi # random theta
noise = -noise_level + np.random.rand(*k_size) * noise_level * 2
# Set COV matrix using Lambdas and Theta
LAMBDA = np.diag([lambda_1, lambda_2])
Q = np.array([[np.cos(theta), -np.sin(theta)],
[np.sin(theta), np.cos(theta)]])
SIGMA = Q @ LAMBDA @ Q.T
INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :]
# Set expectation position (shifting kernel for aligned image)
MU = k_size // 2 - 0.5 * (scale_factor - 1) # - 0.5 * (scale_factor - k_size % 2)
MU = MU[None, None, :, None]
# Create meshgrid for Gaussian
[X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1]))
Z = np.stack([X, Y], 2)[:, :, :, None]
# Calcualte Gaussian for every pixel of the kernel
ZZ = Z - MU
ZZ_t = ZZ.transpose(0, 1, 3, 2)
raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise)
# shift the kernel so it will be centered
# raw_kernel_centered = kernel_shift(raw_kernel, scale_factor)
# Normalize the kernel and return
# kernel = raw_kernel_centered / np.sum(raw_kernel_centered)
kernel = raw_kernel / np.sum(raw_kernel)
return kernel
def fspecial_gaussian(hsize, sigma):
hsize = [hsize, hsize]
siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0]
std = sigma
[x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1))
arg = -(x * x + y * y) / (2 * std * std)
h = np.exp(arg)
h[h < scipy.finfo(float).eps * h.max()] = 0
sumh = h.sum()
if sumh != 0:
h = h / sumh
return h
def fspecial_laplacian(alpha):
alpha = max([0, min([alpha, 1])])
h1 = alpha / (alpha + 1)
h2 = (1 - alpha) / (alpha + 1)
h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]]
h = np.array(h)
return h
def fspecial(filter_type, *args, **kwargs):
'''
python code from:
https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py
'''
if filter_type == 'gaussian':
return fspecial_gaussian(*args, **kwargs)
if filter_type == 'laplacian':
return fspecial_laplacian(*args, **kwargs)
"""
# --------------------------------------------
# degradation models
# --------------------------------------------
"""
def bicubic_degradation(x, sf=3):
'''
Args:
x: HxWxC image, [0, 1]
sf: down-scale factor
Return:
bicubicly downsampled LR image
'''
x = util.imresize_np(x, scale=1 / sf)
return x
def srmd_degradation(x, k, sf=3):
''' blur + bicubic downsampling
Args:
x: HxWxC image, [0, 1]
k: hxw, double
sf: down-scale factor
Return:
downsampled LR image
Reference:
@inproceedings{zhang2018learning,
title={Learning a single convolutional super-resolution network for multiple degradations},
author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
pages={3262--3271},
year={2018}
}
'''
x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap') # 'nearest' | 'mirror'
x = bicubic_degradation(x, sf=sf)
return x
def dpsr_degradation(x, k, sf=3):
''' bicubic downsampling + blur
Args:
x: HxWxC image, [0, 1]
k: hxw, double
sf: down-scale factor
Return:
downsampled LR image
Reference:
@inproceedings{zhang2019deep,
title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels},
author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
pages={1671--1681},
year={2019}
}
'''
x = bicubic_degradation(x, sf=sf)
x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
return x
def classical_degradation(x, k, sf=3):
''' blur + downsampling
Args:
x: HxWxC image, [0, 1]/[0, 255]
k: hxw, double
sf: down-scale factor
Return:
downsampled LR image
'''
x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
# x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2))
st = 0
return x[st::sf, st::sf, ...]
def add_sharpening(img, weight=0.5, radius=50, threshold=10):
"""USM sharpening. borrowed from real-ESRGAN
Input image: I; Blurry image: B.
1. K = I + weight * (I - B)
2. Mask = 1 if abs(I - B) > threshold, else: 0
3. Blur mask:
4. Out = Mask * K + (1 - Mask) * I
Args:
img (Numpy array): Input image, HWC, BGR; float32, [0, 1].
weight (float): Sharp weight. Default: 1.
radius (float): Kernel size of Gaussian blur. Default: 50.
threshold (int):
"""
if radius % 2 == 0:
radius += 1
blur = cv2.GaussianBlur(img, (radius, radius), 0)
residual = img - blur
mask = np.abs(residual) * 255 > threshold
mask = mask.astype('float32')
soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0)
K = img + weight * residual
K = np.clip(K, 0, 1)
return soft_mask * K + (1 - soft_mask) * img
def add_blur(img, sf=4):
wd2 = 4.0 + sf
wd = 2.0 + 0.2 * sf
wd2 = wd2/4
wd = wd/4
if random.random() < 0.5:
l1 = wd2 * random.random()
l2 = wd2 * random.random()
k = anisotropic_Gaussian(ksize=random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2)
else:
k = fspecial('gaussian', random.randint(2, 4) + 3, wd * random.random())
img = ndimage.filters.convolve(img, np.expand_dims(k, axis=2), mode='mirror')
return img
def add_resize(img, sf=4):
rnum = np.random.rand()
if rnum > 0.8: # up
sf1 = random.uniform(1, 2)
elif rnum < 0.7: # down
sf1 = random.uniform(0.5 / sf, 1)
else:
sf1 = 1.0
img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3]))
img = np.clip(img, 0.0, 1.0)
return img
# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
# noise_level = random.randint(noise_level1, noise_level2)
# rnum = np.random.rand()
# if rnum > 0.6: # add color Gaussian noise
# img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
# elif rnum < 0.4: # add grayscale Gaussian noise
# img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
# else: # add noise
# L = noise_level2 / 255.
# D = np.diag(np.random.rand(3))
# U = orth(np.random.rand(3, 3))
# conv = np.dot(np.dot(np.transpose(U), D), U)
# img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
# img = np.clip(img, 0.0, 1.0)
# return img
def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
noise_level = random.randint(noise_level1, noise_level2)
rnum = np.random.rand()
if rnum > 0.6: # add color Gaussian noise
img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
elif rnum < 0.4: # add grayscale Gaussian noise
img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
else: # add noise
L = noise_level2 / 255.
D = np.diag(np.random.rand(3))
U = orth(np.random.rand(3, 3))
conv = np.dot(np.dot(np.transpose(U), D), U)
img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
img = np.clip(img, 0.0, 1.0)
return img
def add_speckle_noise(img, noise_level1=2, noise_level2=25):
noise_level = random.randint(noise_level1, noise_level2)
img = np.clip(img, 0.0, 1.0)
rnum = random.random()
if rnum > 0.6:
img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
elif rnum < 0.4:
img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
else:
L = noise_level2 / 255.
D = np.diag(np.random.rand(3))
U = orth(np.random.rand(3, 3))
conv = np.dot(np.dot(np.transpose(U), D), U)
img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
img = np.clip(img, 0.0, 1.0)
return img
def add_Poisson_noise(img):
img = np.clip((img * 255.0).round(), 0, 255) / 255.
vals = 10 ** (2 * random.random() + 2.0) # [2, 4]
if random.random() < 0.5:
img = np.random.poisson(img * vals).astype(np.float32) / vals
else:
img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114])
img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255.
noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray
img += noise_gray[:, :, np.newaxis]
img = np.clip(img, 0.0, 1.0)
return img
def add_JPEG_noise(img):
quality_factor = random.randint(80, 95)
img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR)
result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor])
img = cv2.imdecode(encimg, 1)
img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB)
return img
def random_crop(lq, hq, sf=4, lq_patchsize=64):
h, w = lq.shape[:2]
rnd_h = random.randint(0, h - lq_patchsize)
rnd_w = random.randint(0, w - lq_patchsize)
lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :]
rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf)
hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :]
return lq, hq
def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
"""
This is the degradation model of BSRGAN from the paper
"Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
----------
img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
sf: scale factor
isp_model: camera ISP model
Returns
-------
img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
"""
isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
sf_ori = sf
h1, w1 = img.shape[:2]
img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop
h, w = img.shape[:2]
if h < lq_patchsize * sf or w < lq_patchsize * sf:
raise ValueError(f'img size ({h1}X{w1}) is too small!')
hq = img.copy()
if sf == 4 and random.random() < scale2_prob: # downsample1
if np.random.rand() < 0.5:
img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])),
interpolation=random.choice([1, 2, 3]))
else:
img = util.imresize_np(img, 1 / 2, True)
img = np.clip(img, 0.0, 1.0)
sf = 2
shuffle_order = random.sample(range(7), 7)
idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
if idx1 > idx2: # keep downsample3 last
shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
for i in shuffle_order:
if i == 0:
img = add_blur(img, sf=sf)
elif i == 1:
img = add_blur(img, sf=sf)
elif i == 2:
a, b = img.shape[1], img.shape[0]
# downsample2
if random.random() < 0.75:
sf1 = random.uniform(1, 2 * sf)
img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])),
interpolation=random.choice([1, 2, 3]))
else:
k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
k_shifted = shift_pixel(k, sf)
k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel
img = ndimage.filters.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror')
img = img[0::sf, 0::sf, ...] # nearest downsampling
img = np.clip(img, 0.0, 1.0)
elif i == 3:
# downsample3
img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
img = np.clip(img, 0.0, 1.0)
elif i == 4:
# add Gaussian noise
img = add_Gaussian_noise(img, noise_level1=2, noise_level2=8)
elif i == 5:
# add JPEG noise
if random.random() < jpeg_prob:
img = add_JPEG_noise(img)
elif i == 6:
# add processed camera sensor noise
if random.random() < isp_prob and isp_model is not None:
with torch.no_grad():
img, hq = isp_model.forward(img.copy(), hq)
# add final JPEG compression noise
img = add_JPEG_noise(img)
# random crop
img, hq = random_crop(img, hq, sf_ori, lq_patchsize)
return img, hq
# todo no isp_model?
def degradation_bsrgan_variant(image, sf=4, isp_model=None):
"""
This is the degradation model of BSRGAN from the paper
"Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
----------
sf: scale factor
isp_model: camera ISP model
Returns
-------
img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
"""
image = util.uint2single(image)
isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
sf_ori = sf
h1, w1 = image.shape[:2]
image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop
h, w = image.shape[:2]
hq = image.copy()
if sf == 4 and random.random() < scale2_prob: # downsample1
if np.random.rand() < 0.5:
image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])),
interpolation=random.choice([1, 2, 3]))
else:
image = util.imresize_np(image, 1 / 2, True)
image = np.clip(image, 0.0, 1.0)
sf = 2
shuffle_order = random.sample(range(7), 7)
idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
if idx1 > idx2: # keep downsample3 last
shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
for i in shuffle_order:
if i == 0:
image = add_blur(image, sf=sf)
# elif i == 1:
# image = add_blur(image, sf=sf)
if i == 0:
pass
elif i == 2:
a, b = image.shape[1], image.shape[0]
# downsample2
if random.random() < 0.8:
sf1 = random.uniform(1, 2 * sf)
image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])),
interpolation=random.choice([1, 2, 3]))
else:
k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
k_shifted = shift_pixel(k, sf)
k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel
image = ndimage.filters.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror')
image = image[0::sf, 0::sf, ...] # nearest downsampling
image = np.clip(image, 0.0, 1.0)
elif i == 3:
# downsample3
image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
image = np.clip(image, 0.0, 1.0)
elif i == 4:
# add Gaussian noise
image = add_Gaussian_noise(image, noise_level1=1, noise_level2=2)
elif i == 5:
# add JPEG noise
if random.random() < jpeg_prob:
image = add_JPEG_noise(image)
#
# elif i == 6:
# # add processed camera sensor noise
# if random.random() < isp_prob and isp_model is not None:
# with torch.no_grad():
# img, hq = isp_model.forward(img.copy(), hq)
# add final JPEG compression noise
image = add_JPEG_noise(image)
image = util.single2uint(image)
example = {"image": image}
return example
if __name__ == '__main__':
print("hey")
img = util.imread_uint('utils/test.png', 3)
img = img[:448, :448]
h = img.shape[0] // 4
print("resizing to", h)
sf = 4
deg_fn = partial(degradation_bsrgan_variant, sf=sf)
for i in range(20):
print(i)
img_hq = img
img_lq = deg_fn(img)["image"]
img_hq, img_lq = util.uint2single(img_hq), util.uint2single(img_lq)
print(img_lq)
img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img_hq)["image"]
print(img_lq.shape)
print("bicubic", img_lq_bicubic.shape)
print(img_hq.shape)
lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
interpolation=0)
lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic),
(int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
interpolation=0)
img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1)
util.imsave(img_concat, str(i) + '.png')

BIN
ldm/modules/image_degradation/utils/test.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 431 KiB

916
ldm/modules/image_degradation/utils_image.py Normal file
View File

@ -0,0 +1,916 @@
import os
import math
import random
import numpy as np
import torch
import cv2
from torchvision.utils import make_grid
from datetime import datetime
#import matplotlib.pyplot as plt # TODO: check with Dominik, also bsrgan.py vs bsrgan_light.py
os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE"
'''
# --------------------------------------------
# Kai Zhang (github: https://github.com/cszn)
# 03/Mar/2019
# --------------------------------------------
# https://github.com/twhui/SRGAN-pyTorch
# https://github.com/xinntao/BasicSR
# --------------------------------------------
'''
IMG_EXTENSIONS = ['.jpg', '.JPG', '.jpeg', '.JPEG', '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP', '.tif']
def is_image_file(filename):
return any(filename.endswith(extension) for extension in IMG_EXTENSIONS)
def get_timestamp():
return datetime.now().strftime('%y%m%d-%H%M%S')
def imshow(x, title=None, cbar=False, figsize=None):
plt.figure(figsize=figsize)
plt.imshow(np.squeeze(x), interpolation='nearest', cmap='gray')
if title:
plt.title(title)
if cbar:
plt.colorbar()
plt.show()
def surf(Z, cmap='rainbow', figsize=None):
plt.figure(figsize=figsize)
ax3 = plt.axes(projection='3d')
w, h = Z.shape[:2]
xx = np.arange(0,w,1)
yy = np.arange(0,h,1)
X, Y = np.meshgrid(xx, yy)
ax3.plot_surface(X,Y,Z,cmap=cmap)
#ax3.contour(X,Y,Z, zdim='z',offset=-2，cmap=cmap)
plt.show()
'''
# --------------------------------------------
# get image pathes
# --------------------------------------------
'''
def get_image_paths(dataroot):
paths = None # return None if dataroot is None
if dataroot is not None:
paths = sorted(_get_paths_from_images(dataroot))
return paths
def _get_paths_from_images(path):
assert os.path.isdir(path), '{:s} is not a valid directory'.format(path)
images = []
for dirpath, _, fnames in sorted(os.walk(path)):
for fname in sorted(fnames):
if is_image_file(fname):
img_path = os.path.join(dirpath, fname)
images.append(img_path)
assert images, '{:s} has no valid image file'.format(path)
return images
'''
# --------------------------------------------
# split large images into small images
# --------------------------------------------
'''
def patches_from_image(img, p_size=512, p_overlap=64, p_max=800):
w, h = img.shape[:2]
patches = []
if w > p_max and h > p_max:
w1 = list(np.arange(0, w-p_size, p_size-p_overlap, dtype=np.int))
h1 = list(np.arange(0, h-p_size, p_size-p_overlap, dtype=np.int))
w1.append(w-p_size)
h1.append(h-p_size)
# print(w1)
# print(h1)
for i in w1:
for j in h1:
patches.append(img[i:i+p_size, j:j+p_size,:])
else:
patches.append(img)
return patches
def imssave(imgs, img_path):
"""
imgs: list, N images of size WxHxC
"""
img_name, ext = os.path.splitext(os.path.basename(img_path))
for i, img in enumerate(imgs):
if img.ndim == 3:
img = img[:, :, [2, 1, 0]]
new_path = os.path.join(os.path.dirname(img_path), img_name+str('_s{:04d}'.format(i))+'.png')
cv2.imwrite(new_path, img)
def split_imageset(original_dataroot, taget_dataroot, n_channels=3, p_size=800, p_overlap=96, p_max=1000):
"""
split the large images from original_dataroot into small overlapped images with size (p_size)x(p_size),
and save them into taget_dataroot; only the images with larger size than (p_max)x(p_max)
will be splitted.
Args:
original_dataroot:
taget_dataroot:
p_size: size of small images
p_overlap: patch size in training is a good choice
p_max: images with smaller size than (p_max)x(p_max) keep unchanged.
"""
paths = get_image_paths(original_dataroot)
for img_path in paths:
# img_name, ext = os.path.splitext(os.path.basename(img_path))
img = imread_uint(img_path, n_channels=n_channels)
patches = patches_from_image(img, p_size, p_overlap, p_max)
imssave(patches, os.path.join(taget_dataroot,os.path.basename(img_path)))
#if original_dataroot == taget_dataroot:
#del img_path
'''
# --------------------------------------------
# makedir
# --------------------------------------------
'''
def mkdir(path):
if not os.path.exists(path):
os.makedirs(path)
def mkdirs(paths):
if isinstance(paths, str):
mkdir(paths)
else:
for path in paths:
mkdir(path)
def mkdir_and_rename(path):
if os.path.exists(path):
new_name = path + '_archived_' + get_timestamp()
print('Path already exists. Rename it to [{:s}]'.format(new_name))
os.rename(path, new_name)
os.makedirs(path)
'''
# --------------------------------------------
# read image from path
# opencv is fast, but read BGR numpy image
# --------------------------------------------
'''
# --------------------------------------------
# get uint8 image of size HxWxn_channles (RGB)
# --------------------------------------------
def imread_uint(path, n_channels=3):
# input: path
# output: HxWx3(RGB or GGG), or HxWx1 (G)
if n_channels == 1:
img = cv2.imread(path, 0) # cv2.IMREAD_GRAYSCALE
img = np.expand_dims(img, axis=2) # HxWx1
elif n_channels == 3:
img = cv2.imread(path, cv2.IMREAD_UNCHANGED) # BGR or G
if img.ndim == 2:
img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB) # GGG
else:
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # RGB
return img
# --------------------------------------------
# matlab's imwrite
# --------------------------------------------
def imsave(img, img_path):
img = np.squeeze(img)
if img.ndim == 3:
img = img[:, :, [2, 1, 0]]
cv2.imwrite(img_path, img)
def imwrite(img, img_path):
img = np.squeeze(img)
if img.ndim == 3:
img = img[:, :, [2, 1, 0]]
cv2.imwrite(img_path, img)
# --------------------------------------------
# get single image of size HxWxn_channles (BGR)
# --------------------------------------------
def read_img(path):
# read image by cv2
# return: Numpy float32, HWC, BGR, [0,1]
img = cv2.imread(path, cv2.IMREAD_UNCHANGED) # cv2.IMREAD_GRAYSCALE
img = img.astype(np.float32) / 255.
if img.ndim == 2:
img = np.expand_dims(img, axis=2)
# some images have 4 channels
if img.shape[2] > 3:
img = img[:, :, :3]
return img
'''
# --------------------------------------------
# image format conversion
# --------------------------------------------
# numpy(single) <---> numpy(unit)
# numpy(single) <---> tensor
# numpy(unit) <---> tensor
# --------------------------------------------
'''
# --------------------------------------------
# numpy(single) [0, 1] <---> numpy(unit)
# --------------------------------------------
def uint2single(img):
return np.float32(img/255.)
def single2uint(img):
return np.uint8((img.clip(0, 1)*255.).round())
def uint162single(img):
return np.float32(img/65535.)
def single2uint16(img):
return np.uint16((img.clip(0, 1)*65535.).round())
# --------------------------------------------
# numpy(unit) (HxWxC or HxW) <---> tensor
# --------------------------------------------
# convert uint to 4-dimensional torch tensor
def uint2tensor4(img):
if img.ndim == 2:
img = np.expand_dims(img, axis=2)
return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.).unsqueeze(0)
# convert uint to 3-dimensional torch tensor
def uint2tensor3(img):
if img.ndim == 2:
img = np.expand_dims(img, axis=2)
return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.)
# convert 2/3/4-dimensional torch tensor to uint
def tensor2uint(img):
img = img.data.squeeze().float().clamp_(0, 1).cpu().numpy()
if img.ndim == 3:
img = np.transpose(img, (1, 2, 0))
return np.uint8((img*255.0).round())
# --------------------------------------------
# numpy(single) (HxWxC) <---> tensor
# --------------------------------------------
# convert single (HxWxC) to 3-dimensional torch tensor
def single2tensor3(img):
return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float()
# convert single (HxWxC) to 4-dimensional torch tensor
def single2tensor4(img):
return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().unsqueeze(0)
# convert torch tensor to single
def tensor2single(img):
img = img.data.squeeze().float().cpu().numpy()
if img.ndim == 3:
img = np.transpose(img, (1, 2, 0))
return img
# convert torch tensor to single
def tensor2single3(img):
img = img.data.squeeze().float().cpu().numpy()
if img.ndim == 3:
img = np.transpose(img, (1, 2, 0))
elif img.ndim == 2:
img = np.expand_dims(img, axis=2)
return img
def single2tensor5(img):
return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float().unsqueeze(0)
def single32tensor5(img):
return torch.from_numpy(np.ascontiguousarray(img)).float().unsqueeze(0).unsqueeze(0)
def single42tensor4(img):
return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float()
# from skimage.io import imread, imsave
def tensor2img(tensor, out_type=np.uint8, min_max=(0, 1)):
'''
Converts a torch Tensor into an image Numpy array of BGR channel order
Input: 4D(B,(3/1),H,W), 3D(C,H,W), or 2D(H,W), any range, RGB channel order
Output: 3D(H,W,C) or 2D(H,W), [0,255], np.uint8 (default)
'''
tensor = tensor.squeeze().float().cpu().clamp_(*min_max) # squeeze first, then clamp
tensor = (tensor - min_max[0]) / (min_max[1] - min_max[0]) # to range [0,1]
n_dim = tensor.dim()
if n_dim == 4:
n_img = len(tensor)
img_np = make_grid(tensor, nrow=int(math.sqrt(n_img)), normalize=False).numpy()
img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0)) # HWC, BGR
elif n_dim == 3:
img_np = tensor.numpy()
img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0)) # HWC, BGR
elif n_dim == 2:
img_np = tensor.numpy()
else:
raise TypeError(
'Only support 4D, 3D and 2D tensor. But received with dimension: {:d}'.format(n_dim))
if out_type == np.uint8:
img_np = (img_np * 255.0).round()
# Important. Unlike matlab, numpy.unit8() WILL NOT round by default.
return img_np.astype(out_type)
'''
# --------------------------------------------
# Augmentation, flipe and/or rotate
# --------------------------------------------
# The following two are enough.
# (1) augmet_img: numpy image of WxHxC or WxH
# (2) augment_img_tensor4: tensor image 1xCxWxH
# --------------------------------------------
'''
def augment_img(img, mode=0):
'''Kai Zhang (github: https://github.com/cszn)
'''
if mode == 0:
return img
elif mode == 1:
return np.flipud(np.rot90(img))
elif mode == 2:
return np.flipud(img)
elif mode == 3:
return np.rot90(img, k=3)
elif mode == 4:
return np.flipud(np.rot90(img, k=2))
elif mode == 5:
return np.rot90(img)
elif mode == 6:
return np.rot90(img, k=2)
elif mode == 7:
return np.flipud(np.rot90(img, k=3))
def augment_img_tensor4(img, mode=0):
'''Kai Zhang (github: https://github.com/cszn)
'''
if mode == 0:
return img
elif mode == 1:
return img.rot90(1, [2, 3]).flip([2])
elif mode == 2:
return img.flip([2])
elif mode == 3:
return img.rot90(3, [2, 3])
elif mode == 4:
return img.rot90(2, [2, 3]).flip([2])
elif mode == 5:
return img.rot90(1, [2, 3])
elif mode == 6:
return img.rot90(2, [2, 3])
elif mode == 7:
return img.rot90(3, [2, 3]).flip([2])
def augment_img_tensor(img, mode=0):
'''Kai Zhang (github: https://github.com/cszn)
'''
img_size = img.size()
img_np = img.data.cpu().numpy()
if len(img_size) == 3:
img_np = np.transpose(img_np, (1, 2, 0))
elif len(img_size) == 4:
img_np = np.transpose(img_np, (2, 3, 1, 0))
img_np = augment_img(img_np, mode=mode)
img_tensor = torch.from_numpy(np.ascontiguousarray(img_np))
if len(img_size) == 3:
img_tensor = img_tensor.permute(2, 0, 1)
elif len(img_size) == 4:
img_tensor = img_tensor.permute(3, 2, 0, 1)
return img_tensor.type_as(img)
def augment_img_np3(img, mode=0):
if mode == 0:
return img
elif mode == 1:
return img.transpose(1, 0, 2)
elif mode == 2:
return img[::-1, :, :]
elif mode == 3:
img = img[::-1, :, :]
img = img.transpose(1, 0, 2)
return img
elif mode == 4:
return img[:, ::-1, :]
elif mode == 5:
img = img[:, ::-1, :]
img = img.transpose(1, 0, 2)
return img
elif mode == 6:
img = img[:, ::-1, :]
img = img[::-1, :, :]
return img
elif mode == 7:
img = img[:, ::-1, :]
img = img[::-1, :, :]
img = img.transpose(1, 0, 2)
return img
def augment_imgs(img_list, hflip=True, rot=True):
# horizontal flip OR rotate
hflip = hflip and random.random() < 0.5
vflip = rot and random.random() < 0.5
rot90 = rot and random.random() < 0.5
def _augment(img):
if hflip:
img = img[:, ::-1, :]
if vflip:
img = img[::-1, :, :]
if rot90:
img = img.transpose(1, 0, 2)
return img
return [_augment(img) for img in img_list]
'''
# --------------------------------------------
# modcrop and shave
# --------------------------------------------
'''
def modcrop(img_in, scale):
# img_in: Numpy, HWC or HW
img = np.copy(img_in)
if img.ndim == 2:
H, W = img.shape
H_r, W_r = H % scale, W % scale
img = img[:H - H_r, :W - W_r]
elif img.ndim == 3:
H, W, C = img.shape
H_r, W_r = H % scale, W % scale
img = img[:H - H_r, :W - W_r, :]
else:
raise ValueError('Wrong img ndim: [{:d}].'.format(img.ndim))
return img
def shave(img_in, border=0):
# img_in: Numpy, HWC or HW
img = np.copy(img_in)
h, w = img.shape[:2]
img = img[border:h-border, border:w-border]
return img
'''
# --------------------------------------------
# image processing process on numpy image
# channel_convert(in_c, tar_type, img_list):
# rgb2ycbcr(img, only_y=True):
# bgr2ycbcr(img, only_y=True):
# ycbcr2rgb(img):
# --------------------------------------------
'''
def rgb2ycbcr(img, only_y=True):
'''same as matlab rgb2ycbcr
only_y: only return Y channel
Input:
uint8, [0, 255]
float, [0, 1]
'''
in_img_type = img.dtype
img.astype(np.float32)
if in_img_type != np.uint8:
img *= 255.
# convert
if only_y:
rlt = np.dot(img, [65.481, 128.553, 24.966]) / 255.0 + 16.0
else:
rlt = np.matmul(img, [[65.481, -37.797, 112.0], [128.553, -74.203, -93.786],
[24.966, 112.0, -18.214]]) / 255.0 + [16, 128, 128]
if in_img_type == np.uint8:
rlt = rlt.round()
else:
rlt /= 255.
return rlt.astype(in_img_type)
def ycbcr2rgb(img):
'''same as matlab ycbcr2rgb
Input:
uint8, [0, 255]
float, [0, 1]
'''
in_img_type = img.dtype
img.astype(np.float32)
if in_img_type != np.uint8:
img *= 255.
# convert
rlt = np.matmul(img, [[0.00456621, 0.00456621, 0.00456621], [0, -0.00153632, 0.00791071],
[0.00625893, -0.00318811, 0]]) * 255.0 + [-222.921, 135.576, -276.836]
if in_img_type == np.uint8:
rlt = rlt.round()
else:
rlt /= 255.
return rlt.astype(in_img_type)
def bgr2ycbcr(img, only_y=True):
'''bgr version of rgb2ycbcr
only_y: only return Y channel
Input:
uint8, [0, 255]
float, [0, 1]
'''
in_img_type = img.dtype
img.astype(np.float32)
if in_img_type != np.uint8:
img *= 255.
# convert
if only_y:
rlt = np.dot(img, [24.966, 128.553, 65.481]) / 255.0 + 16.0
else:
rlt = np.matmul(img, [[24.966, 112.0, -18.214], [128.553, -74.203, -93.786],
[65.481, -37.797, 112.0]]) / 255.0 + [16, 128, 128]
if in_img_type == np.uint8:
rlt = rlt.round()
else:
rlt /= 255.
return rlt.astype(in_img_type)
def channel_convert(in_c, tar_type, img_list):
# conversion among BGR, gray and y
if in_c == 3 and tar_type == 'gray': # BGR to gray
gray_list = [cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) for img in img_list]
return [np.expand_dims(img, axis=2) for img in gray_list]
elif in_c == 3 and tar_type == 'y': # BGR to y
y_list = [bgr2ycbcr(img, only_y=True) for img in img_list]
return [np.expand_dims(img, axis=2) for img in y_list]
elif in_c == 1 and tar_type == 'RGB': # gray/y to BGR
return [cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) for img in img_list]
else:
return img_list
'''
# --------------------------------------------
# metric, PSNR and SSIM
# --------------------------------------------
'''
# --------------------------------------------
# PSNR
# --------------------------------------------
def calculate_psnr(img1, img2, border=0):
# img1 and img2 have range [0, 255]
#img1 = img1.squeeze()
#img2 = img2.squeeze()
if not img1.shape == img2.shape:
raise ValueError('Input images must have the same dimensions.')
h, w = img1.shape[:2]
img1 = img1[border:h-border, border:w-border]
img2 = img2[border:h-border, border:w-border]
img1 = img1.astype(np.float64)
img2 = img2.astype(np.float64)
mse = np.mean((img1 - img2)**2)
if mse == 0:
return float('inf')
return 20 * math.log10(255.0 / math.sqrt(mse))
# --------------------------------------------
# SSIM
# --------------------------------------------
def calculate_ssim(img1, img2, border=0):
'''calculate SSIM
the same outputs as MATLAB's
img1, img2: [0, 255]
'''
#img1 = img1.squeeze()
#img2 = img2.squeeze()
if not img1.shape == img2.shape:
raise ValueError('Input images must have the same dimensions.')
h, w = img1.shape[:2]
img1 = img1[border:h-border, border:w-border]
img2 = img2[border:h-border, border:w-border]
if img1.ndim == 2:
return ssim(img1, img2)
elif img1.ndim == 3:
if img1.shape[2] == 3:
ssims = []
for i in range(3):
ssims.append(ssim(img1[:,:,i], img2[:,:,i]))
return np.array(ssims).mean()
elif img1.shape[2] == 1:
return ssim(np.squeeze(img1), np.squeeze(img2))
else:
raise ValueError('Wrong input image dimensions.')
def ssim(img1, img2):
C1 = (0.01 * 255)**2
C2 = (0.03 * 255)**2
img1 = img1.astype(np.float64)
img2 = img2.astype(np.float64)
kernel = cv2.getGaussianKernel(11, 1.5)
window = np.outer(kernel, kernel.transpose())
mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5] # valid
mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5]
mu1_sq = mu1**2
mu2_sq = mu2**2
mu1_mu2 = mu1 * mu2
sigma1_sq = cv2.filter2D(img1**2, -1, window)[5:-5, 5:-5] - mu1_sq
sigma2_sq = cv2.filter2D(img2**2, -1, window)[5:-5, 5:-5] - mu2_sq
sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2
ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) *
(sigma1_sq + sigma2_sq + C2))
return ssim_map.mean()
'''
# --------------------------------------------
# matlab's bicubic imresize (numpy and torch) [0, 1]
# --------------------------------------------
'''
# matlab 'imresize' function, now only support 'bicubic'
def cubic(x):
absx = torch.abs(x)
absx2 = absx**2
absx3 = absx**3
return (1.5*absx3 - 2.5*absx2 + 1) * ((absx <= 1).type_as(absx)) + \
(-0.5*absx3 + 2.5*absx2 - 4*absx + 2) * (((absx > 1)*(absx <= 2)).type_as(absx))
def calculate_weights_indices(in_length, out_length, scale, kernel, kernel_width, antialiasing):
if (scale < 1) and (antialiasing):
# Use a modified kernel to simultaneously interpolate and antialias- larger kernel width
kernel_width = kernel_width / scale
# Output-space coordinates
x = torch.linspace(1, out_length, out_length)
# Input-space coordinates. Calculate the inverse mapping such that 0.5
# in output space maps to 0.5 in input space, and 0.5+scale in output
# space maps to 1.5 in input space.
u = x / scale + 0.5 * (1 - 1 / scale)
# What is the left-most pixel that can be involved in the computation?
left = torch.floor(u - kernel_width / 2)
# What is the maximum number of pixels that can be involved in the
# computation? Note: it's OK to use an extra pixel here; if the
# corresponding weights are all zero, it will be eliminated at the end
# of this function.
P = math.ceil(kernel_width) + 2
# The indices of the input pixels involved in computing the k-th output
# pixel are in row k of the indices matrix.
indices = left.view(out_length, 1).expand(out_length, P) + torch.linspace(0, P - 1, P).view(
1, P).expand(out_length, P)
# The weights used to compute the k-th output pixel are in row k of the
# weights matrix.
distance_to_center = u.view(out_length, 1).expand(out_length, P) - indices
# apply cubic kernel
if (scale < 1) and (antialiasing):
weights = scale * cubic(distance_to_center * scale)
else:
weights = cubic(distance_to_center)
# Normalize the weights matrix so that each row sums to 1.
weights_sum = torch.sum(weights, 1).view(out_length, 1)
weights = weights / weights_sum.expand(out_length, P)
# If a column in weights is all zero, get rid of it. only consider the first and last column.
weights_zero_tmp = torch.sum((weights == 0), 0)
if not math.isclose(weights_zero_tmp[0], 0, rel_tol=1e-6):
indices = indices.narrow(1, 1, P - 2)
weights = weights.narrow(1, 1, P - 2)
if not math.isclose(weights_zero_tmp[-1], 0, rel_tol=1e-6):
indices = indices.narrow(1, 0, P - 2)
weights = weights.narrow(1, 0, P - 2)
weights = weights.contiguous()
indices = indices.contiguous()
sym_len_s = -indices.min() + 1
sym_len_e = indices.max() - in_length
indices = indices + sym_len_s - 1
return weights, indices, int(sym_len_s), int(sym_len_e)
# --------------------------------------------
# imresize for tensor image [0, 1]
# --------------------------------------------
def imresize(img, scale, antialiasing=True):
# Now the scale should be the same for H and W
# input: img: pytorch tensor, CHW or HW [0,1]
# output: CHW or HW [0,1] w/o round
need_squeeze = True if img.dim() == 2 else False
if need_squeeze:
img.unsqueeze_(0)
in_C, in_H, in_W = img.size()
out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale)
kernel_width = 4
kernel = 'cubic'
# Return the desired dimension order for performing the resize. The
# strategy is to perform the resize first along the dimension with the
# smallest scale factor.
# Now we do not support this.
# get weights and indices
weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices(
in_H, out_H, scale, kernel, kernel_width, antialiasing)
weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices(
in_W, out_W, scale, kernel, kernel_width, antialiasing)
# process H dimension
# symmetric copying
img_aug = torch.FloatTensor(in_C, in_H + sym_len_Hs + sym_len_He, in_W)
img_aug.narrow(1, sym_len_Hs, in_H).copy_(img)
sym_patch = img[:, :sym_len_Hs, :]
inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
sym_patch_inv = sym_patch.index_select(1, inv_idx)
img_aug.narrow(1, 0, sym_len_Hs).copy_(sym_patch_inv)
sym_patch = img[:, -sym_len_He:, :]
inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
sym_patch_inv = sym_patch.index_select(1, inv_idx)
img_aug.narrow(1, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv)
out_1 = torch.FloatTensor(in_C, out_H, in_W)
kernel_width = weights_H.size(1)
for i in range(out_H):
idx = int(indices_H[i][0])
for j in range(out_C):
out_1[j, i, :] = img_aug[j, idx:idx + kernel_width, :].transpose(0, 1).mv(weights_H[i])
# process W dimension
# symmetric copying
out_1_aug = torch.FloatTensor(in_C, out_H, in_W + sym_len_Ws + sym_len_We)
out_1_aug.narrow(2, sym_len_Ws, in_W).copy_(out_1)
sym_patch = out_1[:, :, :sym_len_Ws]
inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long()
sym_patch_inv = sym_patch.index_select(2, inv_idx)
out_1_aug.narrow(2, 0, sym_len_Ws).copy_(sym_patch_inv)
sym_patch = out_1[:, :, -sym_len_We:]
inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long()
sym_patch_inv = sym_patch.index_select(2, inv_idx)
out_1_aug.narrow(2, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv)
out_2 = torch.FloatTensor(in_C, out_H, out_W)
kernel_width = weights_W.size(1)
for i in range(out_W):
idx = int(indices_W[i][0])
for j in range(out_C):
out_2[j, :, i] = out_1_aug[j, :, idx:idx + kernel_width].mv(weights_W[i])
if need_squeeze:
out_2.squeeze_()
return out_2
# --------------------------------------------
# imresize for numpy image [0, 1]
# --------------------------------------------
def imresize_np(img, scale, antialiasing=True):
# Now the scale should be the same for H and W
# input: img: Numpy, HWC or HW [0,1]
# output: HWC or HW [0,1] w/o round
img = torch.from_numpy(img)
need_squeeze = True if img.dim() == 2 else False
if need_squeeze:
img.unsqueeze_(2)
in_H, in_W, in_C = img.size()
out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale)
kernel_width = 4
kernel = 'cubic'
# Return the desired dimension order for performing the resize. The
# strategy is to perform the resize first along the dimension with the
# smallest scale factor.
# Now we do not support this.
# get weights and indices
weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices(
in_H, out_H, scale, kernel, kernel_width, antialiasing)
weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices(
in_W, out_W, scale, kernel, kernel_width, antialiasing)
# process H dimension
# symmetric copying
img_aug = torch.FloatTensor(in_H + sym_len_Hs + sym_len_He, in_W, in_C)
img_aug.narrow(0, sym_len_Hs, in_H).copy_(img)
sym_patch = img[:sym_len_Hs, :, :]
inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long()
sym_patch_inv = sym_patch.index_select(0, inv_idx)
img_aug.narrow(0, 0, sym_len_Hs).copy_(sym_patch_inv)
sym_patch = img[-sym_len_He:, :, :]
inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long()
sym_patch_inv = sym_patch.index_select(0, inv_idx)
img_aug.narrow(0, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv)
out_1 = torch.FloatTensor(out_H, in_W, in_C)
kernel_width = weights_H.size(1)
for i in range(out_H):
idx = int(indices_H[i][0])
for j in range(out_C):
out_1[i, :, j] = img_aug[idx:idx + kernel_width, :, j].transpose(0, 1).mv(weights_H[i])
# process W dimension
# symmetric copying
out_1_aug = torch.FloatTensor(out_H, in_W + sym_len_Ws + sym_len_We, in_C)
out_1_aug.narrow(1, sym_len_Ws, in_W).copy_(out_1)
sym_patch = out_1[:, :sym_len_Ws, :]
inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
sym_patch_inv = sym_patch.index_select(1, inv_idx)
out_1_aug.narrow(1, 0, sym_len_Ws).copy_(sym_patch_inv)
sym_patch = out_1[:, -sym_len_We:, :]
inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
sym_patch_inv = sym_patch.index_select(1, inv_idx)
out_1_aug.narrow(1, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv)
out_2 = torch.FloatTensor(out_H, out_W, in_C)
kernel_width = weights_W.size(1)
for i in range(out_W):
idx = int(indices_W[i][0])
for j in range(out_C):
out_2[:, i, j] = out_1_aug[:, idx:idx + kernel_width, j].mv(weights_W[i])
if need_squeeze:
out_2.squeeze_()
return out_2.numpy()
if __name__ == '__main__':
print('---')
# img = imread_uint('test.bmp', 3)
# img = uint2single(img)
# img_bicubic = imresize_np(img, 1/4)

1
ldm/modules/losses/__init__.py Normal file
View File

@ -0,0 +1 @@
from ldm.modules.losses.contperceptual import LPIPSWithDiscriminator

111
ldm/modules/losses/contperceptual.py Normal file
View File

@ -0,0 +1,111 @@
import torch
import torch.nn as nn
from taming.modules.losses.vqperceptual import * # TODO: taming dependency yes/no?
class LPIPSWithDiscriminator(nn.Module):
def __init__(self, disc_start, logvar_init=0.0, kl_weight=1.0, pixelloss_weight=1.0,
disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=1.0,
perceptual_weight=1.0, use_actnorm=False, disc_conditional=False,
disc_loss="hinge"):
super().__init__()
assert disc_loss in ["hinge", "vanilla"]
self.kl_weight = kl_weight
self.pixel_weight = pixelloss_weight
self.perceptual_loss = LPIPS().eval()
self.perceptual_weight = perceptual_weight
# output log variance
self.logvar = nn.Parameter(torch.ones(size=()) * logvar_init)
self.discriminator = NLayerDiscriminator(input_nc=disc_in_channels,
n_layers=disc_num_layers,
use_actnorm=use_actnorm
).apply(weights_init)
self.discriminator_iter_start = disc_start
self.disc_loss = hinge_d_loss if disc_loss == "hinge" else vanilla_d_loss
self.disc_factor = disc_factor
self.discriminator_weight = disc_weight
self.disc_conditional = disc_conditional
def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None):
if last_layer is not None:
nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0]
g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0]
else:
nll_grads = torch.autograd.grad(nll_loss, self.last_layer[0], retain_graph=True)[0]
g_grads = torch.autograd.grad(g_loss, self.last_layer[0], retain_graph=True)[0]
d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4)
d_weight = torch.clamp(d_weight, 0.0, 1e4).detach()
d_weight = d_weight * self.discriminator_weight
return d_weight
def forward(self, inputs, reconstructions, posteriors, optimizer_idx,
global_step, last_layer=None, cond=None, split="train",
weights=None):
rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous())
if self.perceptual_weight > 0:
p_loss = self.perceptual_loss(inputs.contiguous(), reconstructions.contiguous())
rec_loss = rec_loss + self.perceptual_weight * p_loss
nll_loss = rec_loss / torch.exp(self.logvar) + self.logvar
weighted_nll_loss = nll_loss
if weights is not None:
weighted_nll_loss = weights*nll_loss
weighted_nll_loss = torch.sum(weighted_nll_loss) / weighted_nll_loss.shape[0]
nll_loss = torch.sum(nll_loss) / nll_loss.shape[0]
kl_loss = posteriors.kl()
kl_loss = torch.sum(kl_loss) / kl_loss.shape[0]
# now the GAN part
if optimizer_idx == 0:
# generator update
if cond is None:
assert not self.disc_conditional
logits_fake = self.discriminator(reconstructions.contiguous())
else:
assert self.disc_conditional
logits_fake = self.discriminator(torch.cat((reconstructions.contiguous(), cond), dim=1))
g_loss = -torch.mean(logits_fake)
if self.disc_factor > 0.0:
try:
d_weight = self.calculate_adaptive_weight(nll_loss, g_loss, last_layer=last_layer)
except RuntimeError:
assert not self.training
d_weight = torch.tensor(0.0)
else:
d_weight = torch.tensor(0.0)
disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
loss = weighted_nll_loss + self.kl_weight * kl_loss + d_weight * disc_factor * g_loss
log = {"{}/total_loss".format(split): loss.clone().detach().mean(), "{}/logvar".format(split): self.logvar.detach(),
"{}/kl_loss".format(split): kl_loss.detach().mean(), "{}/nll_loss".format(split): nll_loss.detach().mean(),
"{}/rec_loss".format(split): rec_loss.detach().mean(),
"{}/d_weight".format(split): d_weight.detach(),
"{}/disc_factor".format(split): torch.tensor(disc_factor),
"{}/g_loss".format(split): g_loss.detach().mean(),
}
return loss, log
if optimizer_idx == 1:
# second pass for discriminator update
if cond is None:
logits_real = self.discriminator(inputs.contiguous().detach())
logits_fake = self.discriminator(reconstructions.contiguous().detach())
else:
logits_real = self.discriminator(torch.cat((inputs.contiguous().detach(), cond), dim=1))
logits_fake = self.discriminator(torch.cat((reconstructions.contiguous().detach(), cond), dim=1))
disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
d_loss = disc_factor * self.disc_loss(logits_real, logits_fake)
log = {"{}/disc_loss".format(split): d_loss.clone().detach().mean(),
"{}/logits_real".format(split): logits_real.detach().mean(),
"{}/logits_fake".format(split): logits_fake.detach().mean()
}
return d_loss, log

167
ldm/modules/losses/vqperceptual.py Normal file
View File

@ -0,0 +1,167 @@
import torch
from torch import nn
import torch.nn.functional as F
from einops import repeat
from taming.modules.discriminator.model import NLayerDiscriminator, weights_init
from taming.modules.losses.lpips import LPIPS
from taming.modules.losses.vqperceptual import hinge_d_loss, vanilla_d_loss
def hinge_d_loss_with_exemplar_weights(logits_real, logits_fake, weights):
assert weights.shape[0] == logits_real.shape[0] == logits_fake.shape[0]
loss_real = torch.mean(F.relu(1. - logits_real), dim=[1,2,3])
loss_fake = torch.mean(F.relu(1. + logits_fake), dim=[1,2,3])
loss_real = (weights * loss_real).sum() / weights.sum()
loss_fake = (weights * loss_fake).sum() / weights.sum()
d_loss = 0.5 * (loss_real + loss_fake)
return d_loss
def adopt_weight(weight, global_step, threshold=0, value=0.):
if global_step < threshold:
weight = value
return weight
def measure_perplexity(predicted_indices, n_embed):
# src: https://github.com/karpathy/deep-vector-quantization/blob/main/model.py
# eval cluster perplexity. when perplexity == num_embeddings then all clusters are used exactly equally
encodings = F.one_hot(predicted_indices, n_embed).float().reshape(-1, n_embed)
avg_probs = encodings.mean(0)
perplexity = (-(avg_probs * torch.log(avg_probs + 1e-10)).sum()).exp()
cluster_use = torch.sum(avg_probs > 0)
return perplexity, cluster_use
def l1(x, y):
return torch.abs(x-y)
def l2(x, y):
return torch.pow((x-y), 2)
class VQLPIPSWithDiscriminator(nn.Module):
def __init__(self, disc_start, codebook_weight=1.0, pixelloss_weight=1.0,
disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=1.0,
perceptual_weight=1.0, use_actnorm=False, disc_conditional=False,
disc_ndf=64, disc_loss="hinge", n_classes=None, perceptual_loss="lpips",
pixel_loss="l1"):
super().__init__()
assert disc_loss in ["hinge", "vanilla"]
assert perceptual_loss in ["lpips", "clips", "dists"]
assert pixel_loss in ["l1", "l2"]
self.codebook_weight = codebook_weight
self.pixel_weight = pixelloss_weight
if perceptual_loss == "lpips":
print(f"{self.__class__.__name__}: Running with LPIPS.")
self.perceptual_loss = LPIPS().eval()
else:
raise ValueError(f"Unknown perceptual loss: >> {perceptual_loss} <<")
self.perceptual_weight = perceptual_weight
if pixel_loss == "l1":
self.pixel_loss = l1
else:
self.pixel_loss = l2
self.discriminator = NLayerDiscriminator(input_nc=disc_in_channels,
n_layers=disc_num_layers,
use_actnorm=use_actnorm,
ndf=disc_ndf
).apply(weights_init)
self.discriminator_iter_start = disc_start
if disc_loss == "hinge":
self.disc_loss = hinge_d_loss
elif disc_loss == "vanilla":
self.disc_loss = vanilla_d_loss
else:
raise ValueError(f"Unknown GAN loss '{disc_loss}'.")
print(f"VQLPIPSWithDiscriminator running with {disc_loss} loss.")
self.disc_factor = disc_factor
self.discriminator_weight = disc_weight
self.disc_conditional = disc_conditional
self.n_classes = n_classes
def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None):
if last_layer is not None:
nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0]
g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0]
else:
nll_grads = torch.autograd.grad(nll_loss, self.last_layer[0], retain_graph=True)[0]
g_grads = torch.autograd.grad(g_loss, self.last_layer[0], retain_graph=True)[0]
d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4)
d_weight = torch.clamp(d_weight, 0.0, 1e4).detach()
d_weight = d_weight * self.discriminator_weight
return d_weight
def forward(self, codebook_loss, inputs, reconstructions, optimizer_idx,
global_step, last_layer=None, cond=None, split="train", predicted_indices=None):
if not exists(codebook_loss):
codebook_loss = torch.tensor([0.]).to(inputs.device)
#rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous())
rec_loss = self.pixel_loss(inputs.contiguous(), reconstructions.contiguous())
if self.perceptual_weight > 0:
p_loss = self.perceptual_loss(inputs.contiguous(), reconstructions.contiguous())
rec_loss = rec_loss + self.perceptual_weight * p_loss
else:
p_loss = torch.tensor([0.0])
nll_loss = rec_loss
#nll_loss = torch.sum(nll_loss) / nll_loss.shape[0]
nll_loss = torch.mean(nll_loss)
# now the GAN part
if optimizer_idx == 0:
# generator update
if cond is None:
assert not self.disc_conditional
logits_fake = self.discriminator(reconstructions.contiguous())
else:
assert self.disc_conditional
logits_fake = self.discriminator(torch.cat((reconstructions.contiguous(), cond), dim=1))
g_loss = -torch.mean(logits_fake)
try:
d_weight = self.calculate_adaptive_weight(nll_loss, g_loss, last_layer=last_layer)
except RuntimeError:
assert not self.training
d_weight = torch.tensor(0.0)
disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
loss = nll_loss + d_weight * disc_factor * g_loss + self.codebook_weight * codebook_loss.mean()
log = {"{}/total_loss".format(split): loss.clone().detach().mean(),
"{}/quant_loss".format(split): codebook_loss.detach().mean(),
"{}/nll_loss".format(split): nll_loss.detach().mean(),
"{}/rec_loss".format(split): rec_loss.detach().mean(),
"{}/p_loss".format(split): p_loss.detach().mean(),
"{}/d_weight".format(split): d_weight.detach(),
"{}/disc_factor".format(split): torch.tensor(disc_factor),
"{}/g_loss".format(split): g_loss.detach().mean(),
}
if predicted_indices is not None:
assert self.n_classes is not None
with torch.no_grad():
perplexity, cluster_usage = measure_perplexity(predicted_indices, self.n_classes)
log[f"{split}/perplexity"] = perplexity
log[f"{split}/cluster_usage"] = cluster_usage
return loss, log
if optimizer_idx == 1:
# second pass for discriminator update
if cond is None:
logits_real = self.discriminator(inputs.contiguous().detach())
logits_fake = self.discriminator(reconstructions.contiguous().detach())
else:
logits_real = self.discriminator(torch.cat((inputs.contiguous().detach(), cond), dim=1))
logits_fake = self.discriminator(torch.cat((reconstructions.contiguous().detach(), cond), dim=1))
disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
d_loss = disc_factor * self.disc_loss(logits_real, logits_fake)
log = {"{}/disc_loss".format(split): d_loss.clone().detach().mean(),
"{}/logits_real".format(split): logits_real.detach().mean(),
"{}/logits_fake".format(split): logits_fake.detach().mean()
}
return d_loss, log

641
ldm/modules/x_transformer.py Normal file
View File

@ -0,0 +1,641 @@
"""shout-out to https://github.com/lucidrains/x-transformers/tree/main/x_transformers"""
import torch
from torch import nn, einsum
import torch.nn.functional as F
from functools import partial
from inspect import isfunction
from collections import namedtuple
from einops import rearrange, repeat, reduce
# constants
DEFAULT_DIM_HEAD = 64
Intermediates = namedtuple('Intermediates', [
'pre_softmax_attn',
'post_softmax_attn'
])
LayerIntermediates = namedtuple('Intermediates', [
'hiddens',
'attn_intermediates'
])
class AbsolutePositionalEmbedding(nn.Module):
def __init__(self, dim, max_seq_len):
super().__init__()
self.emb = nn.Embedding(max_seq_len, dim)
self.init_()
def init_(self):
nn.init.normal_(self.emb.weight, std=0.02)
def forward(self, x):
n = torch.arange(x.shape[1], device=x.device)
return self.emb(n)[None, :, :]
class FixedPositionalEmbedding(nn.Module):
def __init__(self, dim):
super().__init__()
inv_freq = 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim))
self.register_buffer('inv_freq', inv_freq)
def forward(self, x, seq_dim=1, offset=0):
t = torch.arange(x.shape[seq_dim], device=x.device).type_as(self.inv_freq) + offset
sinusoid_inp = torch.einsum('i , j -> i j', t, self.inv_freq)
emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1)
return emb[None, :, :]
# helpers
def exists(val):
return val is not None
def default(val, d):
if exists(val):
return val
return d() if isfunction(d) else d
def always(val):
def inner(*args, **kwargs):
return val
return inner
def not_equals(val):
def inner(x):
return x != val
return inner
def equals(val):
def inner(x):
return x == val
return inner
def max_neg_value(tensor):
return -torch.finfo(tensor.dtype).max
# keyword argument helpers
def pick_and_pop(keys, d):
values = list(map(lambda key: d.pop(key), keys))
return dict(zip(keys, values))
def group_dict_by_key(cond, d):
return_val = [dict(), dict()]
for key in d.keys():
match = bool(cond(key))
ind = int(not match)
return_val[ind][key] = d[key]
return (*return_val,)
def string_begins_with(prefix, str):
return str.startswith(prefix)
def group_by_key_prefix(prefix, d):
return group_dict_by_key(partial(string_begins_with, prefix), d)
def groupby_prefix_and_trim(prefix, d):
kwargs_with_prefix, kwargs = group_dict_by_key(partial(string_begins_with, prefix), d)
kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items())))
return kwargs_without_prefix, kwargs
# classes
class Scale(nn.Module):
def __init__(self, value, fn):
super().__init__()
self.value = value
self.fn = fn
def forward(self, x, **kwargs):
x, *rest = self.fn(x, **kwargs)
return (x * self.value, *rest)
class Rezero(nn.Module):
def __init__(self, fn):
super().__init__()
self.fn = fn
self.g = nn.Parameter(torch.zeros(1))
def forward(self, x, **kwargs):
x, *rest = self.fn(x, **kwargs)
return (x * self.g, *rest)
class ScaleNorm(nn.Module):
def __init__(self, dim, eps=1e-5):
super().__init__()
self.scale = dim ** -0.5
self.eps = eps
self.g = nn.Parameter(torch.ones(1))
def forward(self, x):
norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
return x / norm.clamp(min=self.eps) * self.g
class RMSNorm(nn.Module):
def __init__(self, dim, eps=1e-8):
super().__init__()
self.scale = dim ** -0.5
self.eps = eps
self.g = nn.Parameter(torch.ones(dim))
def forward(self, x):
norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
return x / norm.clamp(min=self.eps) * self.g
class Residual(nn.Module):
def forward(self, x, residual):
return x + residual
class GRUGating(nn.Module):
def __init__(self, dim):
super().__init__()
self.gru = nn.GRUCell(dim, dim)
def forward(self, x, residual):
gated_output = self.gru(
rearrange(x, 'b n d -> (b n) d'),
rearrange(residual, 'b n d -> (b n) d')
)
return gated_output.reshape_as(x)
# feedforward
class GEGLU(nn.Module):
def __init__(self, dim_in, dim_out):
super().__init__()
self.proj = nn.Linear(dim_in, dim_out * 2)
def forward(self, x):
x, gate = self.proj(x).chunk(2, dim=-1)
return x * F.gelu(gate)
class FeedForward(nn.Module):
def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
super().__init__()
inner_dim = int(dim * mult)
dim_out = default(dim_out, dim)
project_in = nn.Sequential(
nn.Linear(dim, inner_dim),
nn.GELU()
) if not glu else GEGLU(dim, inner_dim)
self.net = nn.Sequential(
project_in,
nn.Dropout(dropout),
nn.Linear(inner_dim, dim_out)
)
def forward(self, x):
return self.net(x)
# attention.
class Attention(nn.Module):
def __init__(
self,
dim,
dim_head=DEFAULT_DIM_HEAD,
heads=8,
causal=False,
mask=None,
talking_heads=False,
sparse_topk=None,
use_entmax15=False,
num_mem_kv=0,
dropout=0.,
on_attn=False
):
super().__init__()
if use_entmax15:
raise NotImplementedError("Check out entmax activation instead of softmax activation!")
self.scale = dim_head ** -0.5
self.heads = heads
self.causal = causal
self.mask = mask
inner_dim = dim_head * heads
self.to_q = nn.Linear(dim, inner_dim, bias=False)
self.to_k = nn.Linear(dim, inner_dim, bias=False)
self.to_v = nn.Linear(dim, inner_dim, bias=False)
self.dropout = nn.Dropout(dropout)
# talking heads
self.talking_heads = talking_heads
if talking_heads:
self.pre_softmax_proj = nn.Parameter(torch.randn(heads, heads))
self.post_softmax_proj = nn.Parameter(torch.randn(heads, heads))
# explicit topk sparse attention
self.sparse_topk = sparse_topk
# entmax
#self.attn_fn = entmax15 if use_entmax15 else F.softmax
self.attn_fn = F.softmax
# add memory key / values
self.num_mem_kv = num_mem_kv
if num_mem_kv > 0:
self.mem_k = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
self.mem_v = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
# attention on attention
self.attn_on_attn = on_attn
self.to_out = nn.Sequential(nn.Linear(inner_dim, dim * 2), nn.GLU()) if on_attn else nn.Linear(inner_dim, dim)
def forward(
self,
x,
context=None,
mask=None,
context_mask=None,
rel_pos=None,
sinusoidal_emb=None,
prev_attn=None,
mem=None
):
b, n, _, h, talking_heads, device = *x.shape, self.heads, self.talking_heads, x.device
kv_input = default(context, x)
q_input = x
k_input = kv_input
v_input = kv_input
if exists(mem):
k_input = torch.cat((mem, k_input), dim=-2)
v_input = torch.cat((mem, v_input), dim=-2)
if exists(sinusoidal_emb):
# in shortformer, the query would start at a position offset depending on the past cached memory
offset = k_input.shape[-2] - q_input.shape[-2]
q_input = q_input + sinusoidal_emb(q_input, offset=offset)
k_input = k_input + sinusoidal_emb(k_input)
q = self.to_q(q_input)
k = self.to_k(k_input)
v = self.to_v(v_input)
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), (q, k, v))
input_mask = None
if any(map(exists, (mask, context_mask))):
q_mask = default(mask, lambda: torch.ones((b, n), device=device).bool())
k_mask = q_mask if not exists(context) else context_mask
k_mask = default(k_mask, lambda: torch.ones((b, k.shape[-2]), device=device).bool())
q_mask = rearrange(q_mask, 'b i -> b () i ()')
k_mask = rearrange(k_mask, 'b j -> b () () j')
input_mask = q_mask * k_mask
if self.num_mem_kv > 0:
mem_k, mem_v = map(lambda t: repeat(t, 'h n d -> b h n d', b=b), (self.mem_k, self.mem_v))
k = torch.cat((mem_k, k), dim=-2)
v = torch.cat((mem_v, v), dim=-2)
if exists(input_mask):
input_mask = F.pad(input_mask, (self.num_mem_kv, 0), value=True)
dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale
mask_value = max_neg_value(dots)
if exists(prev_attn):
dots = dots + prev_attn
pre_softmax_attn = dots
if talking_heads:
dots = einsum('b h i j, h k -> b k i j', dots, self.pre_softmax_proj).contiguous()
if exists(rel_pos):
dots = rel_pos(dots)
if exists(input_mask):
dots.masked_fill_(~input_mask, mask_value)
del input_mask
if self.causal:
i, j = dots.shape[-2:]
r = torch.arange(i, device=device)
mask = rearrange(r, 'i -> () () i ()') < rearrange(r, 'j -> () () () j')
mask = F.pad(mask, (j - i, 0), value=False)
dots.masked_fill_(mask, mask_value)
del mask
if exists(self.sparse_topk) and self.sparse_topk < dots.shape[-1]:
top, _ = dots.topk(self.sparse_topk, dim=-1)
vk = top[..., -1].unsqueeze(-1).expand_as(dots)
mask = dots < vk
dots.masked_fill_(mask, mask_value)
del mask
attn = self.attn_fn(dots, dim=-1)
post_softmax_attn = attn
attn = self.dropout(attn)
if talking_heads:
attn = einsum('b h i j, h k -> b k i j', attn, self.post_softmax_proj).contiguous()
out = einsum('b h i j, b h j d -> b h i d', attn, v)
out = rearrange(out, 'b h n d -> b n (h d)')
intermediates = Intermediates(
pre_softmax_attn=pre_softmax_attn,
post_softmax_attn=post_softmax_attn
)
return self.to_out(out), intermediates
class AttentionLayers(nn.Module):
def __init__(
self,
dim,
depth,
heads=8,
causal=False,
cross_attend=False,
only_cross=False,
use_scalenorm=False,
use_rmsnorm=False,
use_rezero=False,
rel_pos_num_buckets=32,
rel_pos_max_distance=128,
position_infused_attn=False,
custom_layers=None,
sandwich_coef=None,
par_ratio=None,
residual_attn=False,
cross_residual_attn=False,
macaron=False,
pre_norm=True,
gate_residual=False,
**kwargs
):
super().__init__()
ff_kwargs, kwargs = groupby_prefix_and_trim('ff_', kwargs)
attn_kwargs, _ = groupby_prefix_and_trim('attn_', kwargs)
dim_head = attn_kwargs.get('dim_head', DEFAULT_DIM_HEAD)
self.dim = dim
self.depth = depth
self.layers = nn.ModuleList([])
self.has_pos_emb = position_infused_attn
self.pia_pos_emb = FixedPositionalEmbedding(dim) if position_infused_attn else None
self.rotary_pos_emb = always(None)
assert rel_pos_num_buckets <= rel_pos_max_distance, 'number of relative position buckets must be less than the relative position max distance'
self.rel_pos = None
self.pre_norm = pre_norm
self.residual_attn = residual_attn
self.cross_residual_attn = cross_residual_attn
norm_class = ScaleNorm if use_scalenorm else nn.LayerNorm
norm_class = RMSNorm if use_rmsnorm else norm_class
norm_fn = partial(norm_class, dim)
norm_fn = nn.Identity if use_rezero else norm_fn
branch_fn = Rezero if use_rezero else None
if cross_attend and not only_cross:
default_block = ('a', 'c', 'f')
elif cross_attend and only_cross:
default_block = ('c', 'f')
else:
default_block = ('a', 'f')
if macaron:
default_block = ('f',) + default_block
if exists(custom_layers):
layer_types = custom_layers
elif exists(par_ratio):
par_depth = depth * len(default_block)
assert 1 < par_ratio <= par_depth, 'par ratio out of range'
default_block = tuple(filter(not_equals('f'), default_block))
par_attn = par_depth // par_ratio
depth_cut = par_depth * 2 // 3 # 2 / 3 attention layer cutoff suggested by PAR paper
par_width = (depth_cut + depth_cut // par_attn) // par_attn
assert len(default_block) <= par_width, 'default block is too large for par_ratio'
par_block = default_block + ('f',) * (par_width - len(default_block))
par_head = par_block * par_attn
layer_types = par_head + ('f',) * (par_depth - len(par_head))
elif exists(sandwich_coef):
assert sandwich_coef > 0 and sandwich_coef <= depth, 'sandwich coefficient should be less than the depth'
layer_types = ('a',) * sandwich_coef + default_block * (depth - sandwich_coef) + ('f',) * sandwich_coef
else:
layer_types = default_block * depth
self.layer_types = layer_types
self.num_attn_layers = len(list(filter(equals('a'), layer_types)))
for layer_type in self.layer_types:
if layer_type == 'a':
layer = Attention(dim, heads=heads, causal=causal, **attn_kwargs)
elif layer_type == 'c':
layer = Attention(dim, heads=heads, **attn_kwargs)
elif layer_type == 'f':
layer = FeedForward(dim, **ff_kwargs)
layer = layer if not macaron else Scale(0.5, layer)
else:
raise Exception(f'invalid layer type {layer_type}')
if isinstance(layer, Attention) and exists(branch_fn):
layer = branch_fn(layer)
if gate_residual:
residual_fn = GRUGating(dim)
else:
residual_fn = Residual()
self.layers.append(nn.ModuleList([
norm_fn(),
layer,
residual_fn
]))
def forward(
self,
x,
context=None,
mask=None,
context_mask=None,
mems=None,
return_hiddens=False
):
hiddens = []
intermediates = []
prev_attn = None
prev_cross_attn = None
mems = mems.copy() if exists(mems) else [None] * self.num_attn_layers
for ind, (layer_type, (norm, block, residual_fn)) in enumerate(zip(self.layer_types, self.layers)):
is_last = ind == (len(self.layers) - 1)
if layer_type == 'a':
hiddens.append(x)
layer_mem = mems.pop(0)
residual = x
if self.pre_norm:
x = norm(x)
if layer_type == 'a':
out, inter = block(x, mask=mask, sinusoidal_emb=self.pia_pos_emb, rel_pos=self.rel_pos,
prev_attn=prev_attn, mem=layer_mem)
elif layer_type == 'c':
out, inter = block(x, context=context, mask=mask, context_mask=context_mask, prev_attn=prev_cross_attn)
elif layer_type == 'f':
out = block(x)
x = residual_fn(out, residual)
if layer_type in ('a', 'c'):
intermediates.append(inter)
if layer_type == 'a' and self.residual_attn:
prev_attn = inter.pre_softmax_attn
elif layer_type == 'c' and self.cross_residual_attn:
prev_cross_attn = inter.pre_softmax_attn
if not self.pre_norm and not is_last:
x = norm(x)
if return_hiddens:
intermediates = LayerIntermediates(
hiddens=hiddens,
attn_intermediates=intermediates
)
return x, intermediates
return x
class Encoder(AttentionLayers):
def __init__(self, **kwargs):
assert 'causal' not in kwargs, 'cannot set causality on encoder'
super().__init__(causal=False, **kwargs)
class TransformerWrapper(nn.Module):
def __init__(
self,
*,
num_tokens,
max_seq_len,
attn_layers,
emb_dim=None,
max_mem_len=0.,
emb_dropout=0.,
num_memory_tokens=None,
tie_embedding=False,
use_pos_emb=True
):
super().__init__()
assert isinstance(attn_layers, AttentionLayers), 'attention layers must be one of Encoder or Decoder'
dim = attn_layers.dim
emb_dim = default(emb_dim, dim)
self.max_seq_len = max_seq_len
self.max_mem_len = max_mem_len
self.num_tokens = num_tokens
self.token_emb = nn.Embedding(num_tokens, emb_dim)
self.pos_emb = AbsolutePositionalEmbedding(emb_dim, max_seq_len) if (
use_pos_emb and not attn_layers.has_pos_emb) else always(0)
self.emb_dropout = nn.Dropout(emb_dropout)
self.project_emb = nn.Linear(emb_dim, dim) if emb_dim != dim else nn.Identity()
self.attn_layers = attn_layers
self.norm = nn.LayerNorm(dim)
self.init_()
self.to_logits = nn.Linear(dim, num_tokens) if not tie_embedding else lambda t: t @ self.token_emb.weight.t()
# memory tokens (like [cls]) from Memory Transformers paper
num_memory_tokens = default(num_memory_tokens, 0)
self.num_memory_tokens = num_memory_tokens
if num_memory_tokens > 0:
self.memory_tokens = nn.Parameter(torch.randn(num_memory_tokens, dim))
# let funnel encoder know number of memory tokens, if specified
if hasattr(attn_layers, 'num_memory_tokens'):
attn_layers.num_memory_tokens = num_memory_tokens
def init_(self):
nn.init.normal_(self.token_emb.weight, std=0.02)
def forward(
self,
x,
return_embeddings=False,
mask=None,
return_mems=False,
return_attn=False,
mems=None,
**kwargs
):
b, n, device, num_mem = *x.shape, x.device, self.num_memory_tokens
x = self.token_emb(x)
x += self.pos_emb(x)
x = self.emb_dropout(x)
x = self.project_emb(x)
if num_mem > 0:
mem = repeat(self.memory_tokens, 'n d -> b n d', b=b)
x = torch.cat((mem, x), dim=1)
# auto-handle masking after appending memory tokens
if exists(mask):
mask = F.pad(mask, (num_mem, 0), value=True)
x, intermediates = self.attn_layers(x, mask=mask, mems=mems, return_hiddens=True, **kwargs)
x = self.norm(x)
mem, x = x[:, :num_mem], x[:, num_mem:]
out = self.to_logits(x) if not return_embeddings else x
if return_mems:
hiddens = intermediates.hiddens
new_mems = list(map(lambda pair: torch.cat(pair, dim=-2), zip(mems, hiddens))) if exists(mems) else hiddens
new_mems = list(map(lambda t: t[..., -self.max_mem_len:, :].detach(), new_mems))
return out, new_mems
if return_attn:
attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates))
return out, attn_maps
return out

203
ldm/util.py Normal file
View File

@ -0,0 +1,203 @@
import importlib
import torch
import numpy as np
from collections import abc
from einops import rearrange
from functools import partial
import multiprocessing as mp
from threading import Thread
from queue import Queue
from inspect import isfunction
from PIL import Image, ImageDraw, ImageFont
def log_txt_as_img(wh, xc, size=10):
# wh a tuple of (width, height)
# xc a list of captions to plot
b = len(xc)
txts = list()
for bi in range(b):
txt = Image.new("RGB", wh, color="white")
draw = ImageDraw.Draw(txt)
font = ImageFont.truetype('data/DejaVuSans.ttf', size=size)
nc = int(40 * (wh[0] / 256))
lines = "\n".join(xc[bi][start:start + nc] for start in range(0, len(xc[bi]), nc))
try:
draw.text((0, 0), lines, fill="black", font=font)
except UnicodeEncodeError:
print("Cant encode string for logging. Skipping.")
txt = np.array(txt).transpose(2, 0, 1) / 127.5 - 1.0
txts.append(txt)
txts = np.stack(txts)
txts = torch.tensor(txts)
return txts
def ismap(x):
if not isinstance(x, torch.Tensor):
return False
return (len(x.shape) == 4) and (x.shape[1] > 3)
def isimage(x):
if not isinstance(x, torch.Tensor):
return False
return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1)
def exists(x):
return x is not None
def default(val, d):
if exists(val):
return val
return d() if isfunction(d) else d
def mean_flat(tensor):
"""
https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/nn.py#L86
Take the mean over all non-batch dimensions.
"""
return tensor.mean(dim=list(range(1, len(tensor.shape))))
def count_params(model, verbose=False):
total_params = sum(p.numel() for p in model.parameters())
if verbose:
print(f"{model.__class__.__name__} has {total_params * 1.e-6:.2f} M params.")
return total_params
def instantiate_from_config(config):
if not "target" in config:
if config == '__is_first_stage__':
return None
elif config == "__is_unconditional__":
return None
raise KeyError("Expected key `target` to instantiate.")
return get_obj_from_str(config["target"])(**config.get("params", dict()))
def get_obj_from_str(string, reload=False):
module, cls = string.rsplit(".", 1)
if reload:
module_imp = importlib.import_module(module)
importlib.reload(module_imp)
return getattr(importlib.import_module(module, package=None), cls)
def _do_parallel_data_prefetch(func, Q, data, idx, idx_to_fn=False):
# create dummy dataset instance
# run prefetching
if idx_to_fn:
res = func(data, worker_id=idx)
else:
res = func(data)
Q.put([idx, res])
Q.put("Done")
def parallel_data_prefetch(
func: callable, data, n_proc, target_data_type="ndarray", cpu_intensive=True, use_worker_id=False
):
# if target_data_type not in ["ndarray", "list"]:
# raise ValueError(
# "Data, which is passed to parallel_data_prefetch has to be either of type list or ndarray."
# )
if isinstance(data, np.ndarray) and target_data_type == "list":
raise ValueError("list expected but function got ndarray.")
elif isinstance(data, abc.Iterable):
if isinstance(data, dict):
print(
f'WARNING:"data" argument passed to parallel_data_prefetch is a dict: Using only its values and disregarding keys.'
)
data = list(data.values())
if target_data_type == "ndarray":
data = np.asarray(data)
else:
data = list(data)
else:
raise TypeError(
f"The data, that shall be processed parallel has to be either an np.ndarray or an Iterable, but is actually {type(data)}."
)
if cpu_intensive:
Q = mp.Queue(1000)
proc = mp.Process
else:
Q = Queue(1000)
proc = Thread
# spawn processes
if target_data_type == "ndarray":
arguments = [
[func, Q, part, i, use_worker_id]
for i, part in enumerate(np.array_split(data, n_proc))
]
else:
step = (
int(len(data) / n_proc + 1)
if len(data) % n_proc != 0
else int(len(data) / n_proc)
)
arguments = [
[func, Q, part, i, use_worker_id]
for i, part in enumerate(
[data[i: i + step] for i in range(0, len(data), step)]
)
]
processes = []
for i in range(n_proc):
p = proc(target=_do_parallel_data_prefetch, args=arguments[i])
processes += [p]
# start processes
print(f"Start prefetching...")
import time
start = time.time()
gather_res = [[] for _ in range(n_proc)]
try:
for p in processes:
p.start()
k = 0
while k < n_proc:
# get result
res = Q.get()
if res == "Done":
k += 1
else:
gather_res[res[0]] = res[1]
except Exception as e:
print("Exception: ", e)
for p in processes:
p.terminate()
raise e
finally:
for p in processes:
p.join()
print(f"Prefetching complete. [{time.time() - start} sec.]")
if target_data_type == 'ndarray':
if not isinstance(gather_res[0], np.ndarray):
return np.concatenate([np.asarray(r) for r in gather_res], axis=0)
# order outputs
return np.concatenate(gather_res, axis=0)
elif target_data_type == 'list':
out = []
for r in gather_res:
out.extend(r)
return out
else:
return gather_res

44
models/first_stage_models/kl-f16/config.yaml Normal file
View File

@ -0,0 +1,44 @@
model:
base_learning_rate: 4.5e-06
target: ldm.models.autoencoder.AutoencoderKL
params:
monitor: val/rec_loss
embed_dim: 16
lossconfig:
target: ldm.modules.losses.LPIPSWithDiscriminator
params:
disc_start: 50001
kl_weight: 1.0e-06
disc_weight: 0.5
ddconfig:
double_z: true
z_channels: 16
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 1
- 2
- 2
- 4
num_res_blocks: 2
attn_resolutions:
- 16
dropout: 0.0
data:
target: main.DataModuleFromConfig
params:
batch_size: 6
wrap: true
train:
target: ldm.data.openimages.FullOpenImagesTrain
params:
size: 384
crop_size: 256
validation:
target: ldm.data.openimages.FullOpenImagesValidation
params:
size: 384
crop_size: 256

46
models/first_stage_models/kl-f32/config.yaml Normal file
View File

@ -0,0 +1,46 @@
model:
base_learning_rate: 4.5e-06
target: ldm.models.autoencoder.AutoencoderKL
params:
monitor: val/rec_loss
embed_dim: 64
lossconfig:
target: ldm.modules.losses.LPIPSWithDiscriminator
params:
disc_start: 50001
kl_weight: 1.0e-06
disc_weight: 0.5
ddconfig:
double_z: true
z_channels: 64
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 1
- 2
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions:
- 16
- 8
dropout: 0.0
data:
target: main.DataModuleFromConfig
params:
batch_size: 6
wrap: true
train:
target: ldm.data.openimages.FullOpenImagesTrain
params:
size: 384
crop_size: 256
validation:
target: ldm.data.openimages.FullOpenImagesValidation
params:
size: 384
crop_size: 256

41
models/first_stage_models/kl-f4/config.yaml Normal file
View File

@ -0,0 +1,41 @@
model:
base_learning_rate: 4.5e-06
target: ldm.models.autoencoder.AutoencoderKL
params:
monitor: val/rec_loss
embed_dim: 3
lossconfig:
target: ldm.modules.losses.LPIPSWithDiscriminator
params:
disc_start: 50001
kl_weight: 1.0e-06
disc_weight: 0.5
ddconfig:
double_z: true
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
data:
target: main.DataModuleFromConfig
params:
batch_size: 10
wrap: true
train:
target: ldm.data.openimages.FullOpenImagesTrain
params:
size: 384
crop_size: 256
validation:
target: ldm.data.openimages.FullOpenImagesValidation
params:
size: 384
crop_size: 256

42
models/first_stage_models/kl-f8/config.yaml Normal file
View File

@ -0,0 +1,42 @@
model:
base_learning_rate: 4.5e-06
target: ldm.models.autoencoder.AutoencoderKL
params:
monitor: val/rec_loss
embed_dim: 4
lossconfig:
target: ldm.modules.losses.LPIPSWithDiscriminator
params:
disc_start: 50001
kl_weight: 1.0e-06
disc_weight: 0.5
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
data:
target: main.DataModuleFromConfig
params:
batch_size: 4
wrap: true
train:
target: ldm.data.openimages.FullOpenImagesTrain
params:
size: 384
crop_size: 256
validation:
target: ldm.data.openimages.FullOpenImagesValidation
params:
size: 384
crop_size: 256

49
models/first_stage_models/vq-f16/config.yaml Normal file
View File

@ -0,0 +1,49 @@
model:
base_learning_rate: 4.5e-06
target: ldm.models.autoencoder.VQModel
params:
embed_dim: 8
n_embed: 16384
ddconfig:
double_z: false
z_channels: 8
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 1
- 2
- 2
- 4
num_res_blocks: 2
attn_resolutions:
- 16
dropout: 0.0
lossconfig:
target: taming.modules.losses.vqperceptual.VQLPIPSWithDiscriminator
params:
disc_conditional: false
disc_in_channels: 3
disc_start: 250001
disc_weight: 0.75
disc_num_layers: 2
codebook_weight: 1.0
data:
target: main.DataModuleFromConfig
params:
batch_size: 14
num_workers: 20
wrap: true
train:
target: ldm.data.openimages.FullOpenImagesTrain
params:
size: 384
crop_size: 256
validation:
target: ldm.data.openimages.FullOpenImagesValidation
params:
size: 384
crop_size: 256

46
models/first_stage_models/vq-f4-noattn/config.yaml Normal file
View File

@ -0,0 +1,46 @@
model:
base_learning_rate: 4.5e-06
target: ldm.models.autoencoder.VQModel
params:
embed_dim: 3
n_embed: 8192
monitor: val/rec_loss
ddconfig:
attn_type: none
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: taming.modules.losses.vqperceptual.VQLPIPSWithDiscriminator
params:
disc_conditional: false
disc_in_channels: 3
disc_start: 11
disc_weight: 0.75
codebook_weight: 1.0
data:
target: main.DataModuleFromConfig
params:
batch_size: 8
num_workers: 12
wrap: true
train:
target: ldm.data.openimages.FullOpenImagesTrain
params:
crop_size: 256
validation:
target: ldm.data.openimages.FullOpenImagesValidation
params:
crop_size: 256

45
models/first_stage_models/vq-f4/config.yaml Normal file
View File

@ -0,0 +1,45 @@
model:
base_learning_rate: 4.5e-06
target: ldm.models.autoencoder.VQModel
params:
embed_dim: 3
n_embed: 8192
monitor: val/rec_loss
ddconfig:
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: taming.modules.losses.vqperceptual.VQLPIPSWithDiscriminator
params:
disc_conditional: false
disc_in_channels: 3
disc_start: 0
disc_weight: 0.75
codebook_weight: 1.0
data:
target: main.DataModuleFromConfig
params:
batch_size: 8
num_workers: 16
wrap: true
train:
target: ldm.data.openimages.FullOpenImagesTrain
params:
crop_size: 256
validation:
target: ldm.data.openimages.FullOpenImagesValidation
params:
crop_size: 256

48
models/first_stage_models/vq-f8-n256/config.yaml Normal file
View File

@ -0,0 +1,48 @@
model:
base_learning_rate: 4.5e-06
target: ldm.models.autoencoder.VQModel
params:
embed_dim: 4
n_embed: 256
monitor: val/rec_loss
ddconfig:
double_z: false
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 2
- 4
num_res_blocks: 2
attn_resolutions:
- 32
dropout: 0.0
lossconfig:
target: taming.modules.losses.vqperceptual.VQLPIPSWithDiscriminator
params:
disc_conditional: false
disc_in_channels: 3
disc_start: 250001
disc_weight: 0.75
codebook_weight: 1.0
data:
target: main.DataModuleFromConfig
params:
batch_size: 10
num_workers: 20
wrap: true
train:
target: ldm.data.openimages.FullOpenImagesTrain
params:
size: 384
crop_size: 256
validation:
target: ldm.data.openimages.FullOpenImagesValidation
params:
size: 384
crop_size: 256

48
models/first_stage_models/vq-f8/config.yaml Normal file
View File

@ -0,0 +1,48 @@
model:
base_learning_rate: 4.5e-06
target: ldm.models.autoencoder.VQModel
params:
embed_dim: 4
n_embed: 16384
monitor: val/rec_loss
ddconfig:
double_z: false
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 2
- 4
num_res_blocks: 2
attn_resolutions:
- 32
dropout: 0.0
lossconfig:
target: taming.modules.losses.vqperceptual.VQLPIPSWithDiscriminator
params:
disc_conditional: false
disc_in_channels: 3
disc_num_layers: 2
disc_start: 1
disc_weight: 0.6
codebook_weight: 1.0
data:
target: main.DataModuleFromConfig
params:
batch_size: 10
num_workers: 20
wrap: true
train:
target: ldm.data.openimages.FullOpenImagesTrain
params:
size: 384
crop_size: 256
validation:
target: ldm.data.openimages.FullOpenImagesValidation
params:
size: 384
crop_size: 256

3
models/ldm/.gitignore vendored Normal file
View File

@ -0,0 +1,3 @@
# Exclude the heavyweight models:
*.ckpt
*.ckpt.*

80
models/ldm/bsr_sr/config.yaml Normal file
View File

@ -0,0 +1,80 @@
model:
base_learning_rate: 1.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0155
log_every_t: 100
timesteps: 1000
loss_type: l2
first_stage_key: image
cond_stage_key: LR_image
image_size: 64
channels: 3
concat_mode: true
cond_stage_trainable: false
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 64
in_channels: 6
out_channels: 3
model_channels: 160
attention_resolutions:
- 16
- 8
num_res_blocks: 2
channel_mult:
- 1
- 2
- 2
- 4
num_head_channels: 32
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 3
n_embed: 8192
monitor: val/rec_loss
ddconfig:
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: torch.nn.Identity
data:
target: main.DataModuleFromConfig
params:
batch_size: 64
wrap: false
num_workers: 12
train:
target: ldm.data.openimages.SuperresOpenImagesAdvancedTrain
params:
size: 256
degradation: bsrgan_light
downscale_f: 4
min_crop_f: 0.5
max_crop_f: 1.0
random_crop: true
validation:
target: ldm.data.openimages.SuperresOpenImagesAdvancedValidation
params:
size: 256
degradation: bsrgan_light
downscale_f: 4
min_crop_f: 0.5
max_crop_f: 1.0
random_crop: true

70
models/ldm/celeba256/config.yaml Normal file
View File

@ -0,0 +1,70 @@
model:
base_learning_rate: 2.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0195
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: image
cond_stage_key: class_label
image_size: 64
channels: 3
cond_stage_trainable: false
concat_mode: false
monitor: val/loss
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 64
in_channels: 3
out_channels: 3
model_channels: 224
attention_resolutions:
- 8
- 4
- 2
num_res_blocks: 2
channel_mult:
- 1
- 2
- 3
- 4
num_head_channels: 32
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 3
n_embed: 8192
ddconfig:
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config: __is_unconditional__
data:
target: main.DataModuleFromConfig
params:
batch_size: 48
num_workers: 5
wrap: false
train:
target: ldm.data.faceshq.CelebAHQTrain
params:
size: 256
validation:
target: ldm.data.faceshq.CelebAHQValidation
params:
size: 256

80
models/ldm/cin256/config.yaml Normal file
View File

@ -0,0 +1,80 @@
model:
base_learning_rate: 1.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0195
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: image
cond_stage_key: class_label
image_size: 32
channels: 4
cond_stage_trainable: true
conditioning_key: crossattn
monitor: val/loss_simple_ema
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 32
in_channels: 4
out_channels: 4
model_channels: 256
attention_resolutions:
- 4
- 2
- 1
num_res_blocks: 2
channel_mult:
- 1
- 2
- 4
num_head_channels: 32
use_spatial_transformer: true
transformer_depth: 1
context_dim: 512
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 4
n_embed: 16384
ddconfig:
double_z: false
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 2
- 4
num_res_blocks: 2
attn_resolutions:
- 32
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.ClassEmbedder
params:
embed_dim: 512
key: class_label
data:
target: main.DataModuleFromConfig
params:
batch_size: 64
num_workers: 12
wrap: false
train:
target: ldm.data.imagenet.ImageNetTrain
params:
config:
size: 256
validation:
target: ldm.data.imagenet.ImageNetValidation
params:
config:
size: 256

70
models/ldm/ffhq256/config.yaml Normal file
View File

@ -0,0 +1,70 @@
model:
base_learning_rate: 2.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0195
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: image
cond_stage_key: class_label
image_size: 64
channels: 3
cond_stage_trainable: false
concat_mode: false
monitor: val/loss
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 64
in_channels: 3
out_channels: 3
model_channels: 224
attention_resolutions:
- 8
- 4
- 2
num_res_blocks: 2
channel_mult:
- 1
- 2
- 3
- 4
num_head_channels: 32
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 3
n_embed: 8192
ddconfig:
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config: __is_unconditional__
data:
target: main.DataModuleFromConfig
params:
batch_size: 42
num_workers: 5
wrap: false
train:
target: ldm.data.faceshq.FFHQTrain
params:
size: 256
validation:
target: ldm.data.faceshq.FFHQValidation
params:
size: 256

67
models/ldm/inpainting_big/config.yaml Normal file
View File

@ -0,0 +1,67 @@
model:
base_learning_rate: 1.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0205
log_every_t: 100
timesteps: 1000
loss_type: l1
first_stage_key: image
cond_stage_key: masked_image
image_size: 64
channels: 3
concat_mode: true
monitor: val/loss
scheduler_config:
target: ldm.lr_scheduler.LambdaWarmUpCosineScheduler
params:
verbosity_interval: 0
warm_up_steps: 1000
max_decay_steps: 50000
lr_start: 0.001
lr_max: 0.1
lr_min: 0.0001
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 64
in_channels: 7
out_channels: 3
model_channels: 256
attention_resolutions:
- 8
- 4
- 2
num_res_blocks: 2
channel_mult:
- 1
- 2
- 3
- 4
num_heads: 8
resblock_updown: true
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 3
n_embed: 8192
monitor: val/rec_loss
ddconfig:
attn_type: none
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: ldm.modules.losses.contperceptual.DummyLoss
cond_stage_config: __is_first_stage__

81
models/ldm/layout2img-openimages256/config.yaml Normal file
View File

@ -0,0 +1,81 @@
model:
base_learning_rate: 2.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0205
log_every_t: 100
timesteps: 1000
loss_type: l1
first_stage_key: image
cond_stage_key: coordinates_bbox
image_size: 64
channels: 3
conditioning_key: crossattn
cond_stage_trainable: true
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 64
in_channels: 3
out_channels: 3
model_channels: 128
attention_resolutions:
- 8
- 4
- 2
num_res_blocks: 2
channel_mult:
- 1
- 2
- 3
- 4
num_head_channels: 32
use_spatial_transformer: true
transformer_depth: 3
context_dim: 512
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 3
n_embed: 8192
monitor: val/rec_loss
ddconfig:
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.BERTEmbedder
params:
n_embed: 512
n_layer: 16
vocab_size: 8192
max_seq_len: 92
use_tokenizer: false
monitor: val/loss_simple_ema
data:
target: main.DataModuleFromConfig
params:
batch_size: 24
wrap: false
num_workers: 10
train:
target: ldm.data.openimages.OpenImagesBBoxTrain
params:
size: 256
validation:
target: ldm.data.openimages.OpenImagesBBoxValidation
params:
size: 256

70
models/ldm/lsun_beds256/config.yaml Normal file
View File

@ -0,0 +1,70 @@
model:
base_learning_rate: 2.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0195
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: image
cond_stage_key: class_label
image_size: 64
channels: 3
cond_stage_trainable: false
concat_mode: false
monitor: val/loss
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 64
in_channels: 3
out_channels: 3
model_channels: 224
attention_resolutions:
- 8
- 4
- 2
num_res_blocks: 2
channel_mult:
- 1
- 2
- 3
- 4
num_head_channels: 32
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 3
n_embed: 8192
ddconfig:
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config: __is_unconditional__
data:
target: main.DataModuleFromConfig
params:
batch_size: 48
num_workers: 5
wrap: false
train:
target: ldm.data.lsun.LSUNBedroomsTrain
params:
size: 256
validation:
target: ldm.data.lsun.LSUNBedroomsValidation
params:
size: 256

92
models/ldm/lsun_churches256/config.yaml Normal file
View File

@ -0,0 +1,92 @@
model:
base_learning_rate: 5.0e-05
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0155
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
loss_type: l1
first_stage_key: image
cond_stage_key: image
image_size: 32
channels: 4
cond_stage_trainable: false
concat_mode: false
scale_by_std: true
monitor: val/loss_simple_ema
scheduler_config:
target: ldm.lr_scheduler.LambdaLinearScheduler
params:
warm_up_steps:
- 10000
cycle_lengths:
- 10000000000000
f_start:
- 1.0e-06
f_max:
- 1.0
f_min:
- 1.0
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 32
in_channels: 4
out_channels: 4
model_channels: 192
attention_resolutions:
- 1
- 2
- 4
- 8
num_res_blocks: 2
channel_mult:
- 1
- 2
- 2
- 4
- 4
num_heads: 8
use_scale_shift_norm: true
resblock_updown: true
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config: '__is_unconditional__'
data:
target: main.DataModuleFromConfig
params:
batch_size: 96
num_workers: 5
wrap: false
train:
target: ldm.data.lsun.LSUNChurchesTrain
params:
size: 256
validation:
target: ldm.data.lsun.LSUNChurchesValidation
params:
size: 256

59
models/ldm/semantic_synthesis256/config.yaml Normal file
View File

@ -0,0 +1,59 @@
model:
base_learning_rate: 1.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0205
log_every_t: 100
timesteps: 1000
loss_type: l1
first_stage_key: image
cond_stage_key: segmentation
image_size: 64
channels: 3
concat_mode: true
cond_stage_trainable: true
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 64
in_channels: 6
out_channels: 3
model_channels: 128
attention_resolutions:
- 32
- 16
- 8
num_res_blocks: 2
channel_mult:
- 1
- 4
- 8
num_heads: 8
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 3
n_embed: 8192
ddconfig:
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.SpatialRescaler
params:
n_stages: 2
in_channels: 182
out_channels: 3

78
models/ldm/semantic_synthesis512/config.yaml Normal file
View File

@ -0,0 +1,78 @@
model:
base_learning_rate: 1.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0205
log_every_t: 100
timesteps: 1000
loss_type: l1
first_stage_key: image
cond_stage_key: segmentation
image_size: 128
channels: 3
concat_mode: true
cond_stage_trainable: true
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 128
in_channels: 6
out_channels: 3
model_channels: 128
attention_resolutions:
- 32
- 16
- 8
num_res_blocks: 2
channel_mult:
- 1
- 4
- 8
num_heads: 8
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 3
n_embed: 8192
monitor: val/rec_loss
ddconfig:
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.SpatialRescaler
params:
n_stages: 2
in_channels: 182
out_channels: 3
data:
target: main.DataModuleFromConfig
params:
batch_size: 8
wrap: false
num_workers: 10
train:
target: ldm.data.landscapes.RFWTrain
params:
size: 768
crop_size: 512
segmentation_to_float32: true
validation:
target: ldm.data.landscapes.RFWValidation
params:
size: 768
crop_size: 512
segmentation_to_float32: true

0
models/ldm/stable-diffusion-v1/add-model-ckpt-here Normal file
View File

77
models/ldm/text2img256/config.yaml Normal file
View File

@ -0,0 +1,77 @@
model:
base_learning_rate: 2.0e-06
target: ldm.models.diffusion.ddpm.LatentDiffusion
params:
linear_start: 0.0015
linear_end: 0.0195
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: image
cond_stage_key: caption
image_size: 64
channels: 3
cond_stage_trainable: true
conditioning_key: crossattn
monitor: val/loss_simple_ema
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 64
in_channels: 3
out_channels: 3
model_channels: 192
attention_resolutions:
- 8
- 4
- 2
num_res_blocks: 2
channel_mult:
- 1
- 2
- 3
- 5
num_head_channels: 32
use_spatial_transformer: true
transformer_depth: 1
context_dim: 640
first_stage_config:
target: ldm.models.autoencoder.VQModelInterface
params:
embed_dim: 3
n_embed: 8192
ddconfig:
double_z: false
z_channels: 3
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.BERTEmbedder
params:
n_embed: 640
n_layer: 32
data:
target: main.DataModuleFromConfig
params:
batch_size: 28
num_workers: 10
wrap: false
train:
target: ldm.data.previews.pytorch_dataset.PreviewsTrain
params:
size: 256
validation:
target: ldm.data.previews.pytorch_dataset.PreviewsValidation
params:
size: 256

774
optimizedSD/ddpm.py Normal file
View File

@ -0,0 +1,774 @@
"""
wild mixture of
https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
https://github.com/CompVis/taming-transformers
-- merci
"""
import time
import torch
from einops import rearrange
from tqdm import tqdm
from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
from ldm.models.autoencoder import VQModelInterface
import torch.nn as nn
import numpy as np
import pytorch_lightning as pl
from functools import partial
from pytorch_lightning.utilities.distributed import rank_zero_only
from ldm.util import exists, default, instantiate_from_config
from ldm.modules.diffusionmodules.util import make_beta_schedule
from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like
from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
def disabled_train(self):
"""Overwrite model.train with this function to make sure train/eval mode
does not change anymore."""
return self
class DDPM(pl.LightningModule):
# classic DDPM with Gaussian diffusion, in image space
def __init__(self,
timesteps=1000,
beta_schedule="linear",
ckpt_path=None,
ignore_keys=[],
load_only_unet=False,
monitor="val/loss",
use_ema=True,
first_stage_key="image",
image_size=256,
channels=3,
log_every_t=100,
clip_denoised=True,
linear_start=1e-4,
linear_end=2e-2,
cosine_s=8e-3,
given_betas=None,
original_elbo_weight=0.,
v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
l_simple_weight=1.,
conditioning_key=None,
parameterization="eps", # all assuming fixed variance schedules
scheduler_config=None,
use_positional_encodings=False,
):
super().__init__()
assert parameterization in ["eps", "x0"], 'currently only supporting "eps" and "x0"'
self.parameterization = parameterization
print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
self.cond_stage_model = None
self.clip_denoised = clip_denoised
self.log_every_t = log_every_t
self.first_stage_key = first_stage_key
self.image_size = image_size # try conv?
self.channels = channels
self.use_positional_encodings = use_positional_encodings
self.use_scheduler = scheduler_config is not None
if self.use_scheduler:
self.scheduler_config = scheduler_config
self.v_posterior = v_posterior
self.original_elbo_weight = original_elbo_weight
self.l_simple_weight = l_simple_weight
if monitor is not None:
self.monitor = monitor
if ckpt_path is not None:
self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet)
self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
if exists(given_betas):
betas = given_betas
else:
betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
cosine_s=cosine_s)
alphas = 1. - betas
alphas_cumprod = np.cumprod(alphas, axis=0)
alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
timesteps, = betas.shape
self.num_timesteps = int(timesteps)
self.linear_start = linear_start
self.linear_end = linear_end
assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
to_torch = partial(torch.tensor, dtype=torch.float32)
self.register_buffer('betas', to_torch(betas))
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
class FirstStage(DDPM):
"""main class"""
def __init__(self,
first_stage_config,
num_timesteps_cond=None,
cond_stage_key="image",
cond_stage_trainable=False,
concat_mode=True,
cond_stage_forward=None,
conditioning_key=None,
scale_factor=1.0,
scale_by_std=False,
*args, **kwargs):
self.num_timesteps_cond = default(num_timesteps_cond, 1)
self.scale_by_std = scale_by_std
assert self.num_timesteps_cond <= kwargs['timesteps']
# for backwards compatibility after implementation of DiffusionWrapper
if conditioning_key is None:
conditioning_key = 'concat' if concat_mode else 'crossattn'
ckpt_path = kwargs.pop("ckpt_path", None)
ignore_keys = kwargs.pop("ignore_keys", [])
super().__init__()
self.concat_mode = concat_mode
self.cond_stage_trainable = cond_stage_trainable
self.cond_stage_key = cond_stage_key
try:
self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
except:
self.num_downs = 0
if not scale_by_std:
self.scale_factor = scale_factor
self.instantiate_first_stage(first_stage_config)
self.cond_stage_forward = cond_stage_forward
self.clip_denoised = False
self.bbox_tokenizer = None
self.restarted_from_ckpt = False
if ckpt_path is not None:
self.init_from_ckpt(ckpt_path, ignore_keys)
self.restarted_from_ckpt = True
def instantiate_first_stage(self, config):
model = instantiate_from_config(config)
self.first_stage_model = model.eval()
self.first_stage_model.train = disabled_train
for param in self.first_stage_model.parameters():
param.requires_grad = False
def get_first_stage_encoding(self, encoder_posterior):
if isinstance(encoder_posterior, DiagonalGaussianDistribution):
z = encoder_posterior.sample()
elif isinstance(encoder_posterior, torch.Tensor):
z = encoder_posterior
else:
raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
return self.scale_factor * z
@torch.no_grad()
def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
if predict_cids:
if z.dim() == 4:
z = torch.argmax(z.exp(), dim=1).long()
z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
z = rearrange(z, 'b h w c -> b c h w').contiguous()
z = 1. / self.scale_factor * z
if hasattr(self, "split_input_params"):
if isinstance(self.first_stage_model, VQModelInterface):
return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
else:
return self.first_stage_model.decode(z)
else:
if isinstance(self.first_stage_model, VQModelInterface):
return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
else:
return self.first_stage_model.decode(z)
@torch.no_grad()
def encode_first_stage(self, x):
if hasattr(self, "split_input_params"):
if self.split_input_params["patch_distributed_vq"]:
ks = self.split_input_params["ks"] # eg. (128, 128)
stride = self.split_input_params["stride"] # eg. (64, 64)
df = self.split_input_params["vqf"]
self.split_input_params['original_image_size'] = x.shape[-2:]
bs, nc, h, w = x.shape
if ks[0] > h or ks[1] > w:
ks = (min(ks[0], h), min(ks[1], w))
print("reducing Kernel")
if stride[0] > h or stride[1] > w:
stride = (min(stride[0], h), min(stride[1], w))
print("reducing stride")
fold, unfold, normalization, weighting = self.get_fold_unfold(x, ks, stride, df=df)
z = unfold(x) # (bn, nc * prod(**ks), L)
# Reshape to img shape
z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
output_list = [self.first_stage_model.encode(z[:, :, :, :, i])
for i in range(z.shape[-1])]
o = torch.stack(output_list, axis=-1)
o = o * weighting
# Reverse reshape to img shape
o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
# stitch crops together
decoded = fold(o)
decoded = decoded / normalization
return decoded
else:
return self.first_stage_model.encode(x)
else:
return self.first_stage_model.encode(x)
class CondStage(DDPM):
"""main class"""
def __init__(self,
cond_stage_config,
num_timesteps_cond=None,
cond_stage_key="image",
cond_stage_trainable=False,
concat_mode=True,
cond_stage_forward=None,
conditioning_key=None,
scale_factor=1.0,
scale_by_std=False,
*args, **kwargs):
self.num_timesteps_cond = default(num_timesteps_cond, 1)
self.scale_by_std = scale_by_std
assert self.num_timesteps_cond <= kwargs['timesteps']
# for backwards compatibility after implementation of DiffusionWrapper
if conditioning_key is None:
conditioning_key = 'concat' if concat_mode else 'crossattn'
if cond_stage_config == '__is_unconditional__':
conditioning_key = None
ckpt_path = kwargs.pop("ckpt_path", None)
ignore_keys = kwargs.pop("ignore_keys", [])
super().__init__()
self.concat_mode = concat_mode
self.cond_stage_trainable = cond_stage_trainable
self.cond_stage_key = cond_stage_key
self.num_downs = 0
if not scale_by_std:
self.scale_factor = scale_factor
self.instantiate_cond_stage(cond_stage_config)
self.cond_stage_forward = cond_stage_forward
self.clip_denoised = False
self.bbox_tokenizer = None
self.restarted_from_ckpt = False
if ckpt_path is not None:
self.init_from_ckpt(ckpt_path, ignore_keys)
self.restarted_from_ckpt = True
def instantiate_cond_stage(self, config):
if not self.cond_stage_trainable:
if config == "__is_first_stage__":
print("Using first stage also as cond stage.")
self.cond_stage_model = self.first_stage_model
elif config == "__is_unconditional__":
print(f"Training {self.__class__.__name__} as an unconditional model.")
self.cond_stage_model = None
# self.be_unconditional = True
else:
model = instantiate_from_config(config)
self.cond_stage_model = model.eval()
self.cond_stage_model.train = disabled_train
for param in self.cond_stage_model.parameters():
param.requires_grad = False
else:
assert config != '__is_first_stage__'
assert config != '__is_unconditional__'
model = instantiate_from_config(config)
self.cond_stage_model = model
def get_learned_conditioning(self, c):
if self.cond_stage_forward is None:
if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
c = self.cond_stage_model.encode(c)
if isinstance(c, DiagonalGaussianDistribution):
c = c.mode()
else:
c = self.cond_stage_model(c)
else:
assert hasattr(self.cond_stage_model, self.cond_stage_forward)
c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
return c
class DiffusionWrapper(pl.LightningModule):
def __init__(self, diff_model_config):
super().__init__()
self.diffusion_model = instantiate_from_config(diff_model_config)
def forward(self, x, t, cc):
out = self.diffusion_model(x, t, context=cc)
return out
class DiffusionWrapperOut(pl.LightningModule):
def __init__(self, diff_model_config):
super().__init__()
self.diffusion_model = instantiate_from_config(diff_model_config)
def forward(self, h,emb,tp,hs, cc):
return self.diffusion_model(h,emb,tp,hs, context=cc)
class UNet(DDPM):
"""main class"""
def __init__(self,
unetConfigEncode,
unetConfigDecode,
num_timesteps_cond=None,
cond_stage_key="image",
cond_stage_trainable=False,
concat_mode=True,
cond_stage_forward=None,
conditioning_key=None,
scale_factor=1.0,
unet_bs = 1,
scale_by_std=False,
*args, **kwargs):
self.num_timesteps_cond = default(num_timesteps_cond, 1)
self.scale_by_std = scale_by_std
assert self.num_timesteps_cond <= kwargs['timesteps']
# for backwards compatibility after implementation of DiffusionWrapper
if conditioning_key is None:
conditioning_key = 'concat' if concat_mode else 'crossattn'
ckpt_path = kwargs.pop("ckpt_path", None)
ignore_keys = kwargs.pop("ignore_keys", [])
super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
self.concat_mode = concat_mode
self.cond_stage_trainable = cond_stage_trainable
self.cond_stage_key = cond_stage_key
self.num_downs = 0
self.cdevice = "cuda"
self.unetConfigEncode = unetConfigEncode
self.unetConfigDecode = unetConfigDecode
if not scale_by_std:
self.scale_factor = scale_factor
else:
self.register_buffer('scale_factor', torch.tensor(scale_factor))
self.cond_stage_forward = cond_stage_forward
self.clip_denoised = False
self.bbox_tokenizer = None
self.model1 = DiffusionWrapper(self.unetConfigEncode)
self.model2 = DiffusionWrapperOut(self.unetConfigDecode)
self.model1.eval()
self.model2.eval()
self.turbo = False
self.unet_bs = unet_bs
self.restarted_from_ckpt = False
if ckpt_path is not None:
self.init_from_ckpt(ckpt_path, ignore_keys)
self.restarted_from_ckpt = True
def make_cond_schedule(self, ):
self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
self.cond_ids[:self.num_timesteps_cond] = ids
@rank_zero_only
@torch.no_grad()
def on_train_batch_start(self, batch, batch_idx):
# only for very first batch
if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
# set rescale weight to 1./std of encodings
print("### USING STD-RESCALING ###")
x = super().get_input(batch, self.first_stage_key)
x = x.to(self.cdevice)
encoder_posterior = self.encode_first_stage(x)
z = self.get_first_stage_encoding(encoder_posterior).detach()
del self.scale_factor
self.register_buffer('scale_factor', 1. / z.flatten().std())
print(f"setting self.scale_factor to {self.scale_factor}")
print("### USING STD-RESCALING ###")
def apply_model(self, x_noisy, t, cond, return_ids=False):
if(not self.turbo):
self.model1.to(self.cdevice)
step = self.unet_bs
h,emb,hs = self.model1(x_noisy[0:step], t[:step], cond[:step])
bs = cond.shape[0]
assert bs%2 == 0
lenhs = len(hs)
for i in range(step,bs,step):
h_temp,emb_temp,hs_temp = self.model1(x_noisy[i:i+step], t[i:i+step], cond[i:i+step])
h = torch.cat((h,h_temp))
emb = torch.cat((emb,emb_temp))
for j in range(lenhs):
hs[j] = torch.cat((hs[j], hs_temp[j]))
if(not self.turbo):
self.model1.to("cpu")
self.model2.to(self.cdevice)
hs_temp = [hs[j][:step] for j in range(lenhs)]
x_recon = self.model2(h[:step],emb[:step],x_noisy.dtype,hs_temp,cond[:step])
for i in range(step,bs,step):
hs_temp = [hs[j][i:i+step] for j in range(lenhs)]
x_recon1 = self.model2(h[i:i+step],emb[i:i+step],x_noisy.dtype,hs_temp,cond[i:i+step])
x_recon = torch.cat((x_recon, x_recon1))
if(not self.turbo):
self.model2.to("cpu")
if isinstance(x_recon, tuple) and not return_ids:
return x_recon[0]
else:
return x_recon
def register_buffer1(self, name, attr):
if type(attr) == torch.Tensor:
if attr.device != torch.device(self.cdevice):
attr = attr.to(torch.device(self.cdevice))
setattr(self, name, attr)
def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
num_ddpm_timesteps=self.num_timesteps,verbose=verbose)
alphas_cumprod = self.alphas_cumprod
assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
to_torch = lambda x: x.to(self.cdevice)
self.register_buffer1('betas', to_torch(self.betas))
self.register_buffer1('alphas_cumprod', to_torch(alphas_cumprod))
self.register_buffer1('alphas_cumprod_prev', to_torch(self.alphas_cumprod_prev))
self.register_buffer1('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
# ddim sampling parameters
ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=self.alphas_cumprod.cpu(),
ddim_timesteps=self.ddim_timesteps,
eta=ddim_eta,verbose=verbose)
self.register_buffer1('ddim_sigmas', ddim_sigmas)
self.register_buffer1('ddim_alphas', ddim_alphas)
self.register_buffer1('ddim_alphas_prev', ddim_alphas_prev)
self.register_buffer1('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
self.ddim_sqrt_one_minus_alphas = np.sqrt(1. - ddim_alphas)
sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
(1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
1 - self.alphas_cumprod / self.alphas_cumprod_prev))
self.register_buffer1('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
@torch.no_grad()
def sample(self,
S,
batch_size,
shape,
seed,
conditioning=None,
callback=None,
img_callback=None,
quantize_x0=False,
eta=0.,
mask=None,
x0=None,
temperature=1.,
noise_dropout=0.,
score_corrector=None,
corrector_kwargs=None,
verbose=True,
x_T=None,
log_every_t=100,
unconditional_guidance_scale=1.,
unconditional_conditioning=None,
):
if conditioning is not None:
if isinstance(conditioning, dict):
cbs = conditioning[list(conditioning.keys())[0]].shape[0]
if cbs != batch_size:
print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
else:
if conditioning.shape[0] != batch_size:
print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=False)
# sampling
C, H, W = shape
size = (batch_size, C, H, W)
print(f'Data shape for PLMS sampling is {size}')
if(self.turbo):
self.model1.to(self.cdevice)
self.model2.to(self.cdevice)
samples = self.plms_sampling(conditioning, size, seed,
callback=callback,
img_callback=img_callback,
quantize_denoised=quantize_x0,
mask=mask, x0=x0,
ddim_use_original_steps=False,
noise_dropout=noise_dropout,
temperature=temperature,
score_corrector=score_corrector,
corrector_kwargs=corrector_kwargs,
x_T=x_T,
log_every_t=log_every_t,
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=unconditional_conditioning,
)
if(self.turbo):
self.model1.to("cpu")
self.model2.to("cpu")
return samples
@torch.no_grad()
def plms_sampling(self, cond, shape, seed,
x_T=None, ddim_use_original_steps=False,
callback=None, timesteps=None, quantize_denoised=False,
mask=None, x0=None, img_callback=None, log_every_t=100,
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
unconditional_guidance_scale=1., unconditional_conditioning=None,):
device = self.betas.device
b = shape[0]
if x_T is None:
_, b1, b2, b3 = shape
img_shape = (1, b1, b2, b3)
tens = []
print("seeds used = ", [seed+s for s in range(b)])
for _ in range(b):
torch.manual_seed(seed)
tens.append(torch.randn(img_shape, device=device))
seed+=1
img = torch.cat(tens)
del tens
else:
img = x_T
if timesteps is None:
timesteps = self.num_timesteps if ddim_use_original_steps else self.ddim_timesteps
elif timesteps is not None and not ddim_use_original_steps:
subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
timesteps = self.ddim_timesteps[:subset_end]
time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps)
total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
print(f"Running PLMS Sampling with {total_steps} timesteps")
iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps)
old_eps = []
for i, step in enumerate(iterator):
index = total_steps - i - 1
ts = torch.full((b,), step, device=device, dtype=torch.long)
ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long)
if mask is not None:
assert x0 is not None
img_orig = self.q_sample(x0, ts) # TODO: deterministic forward pass?
img = img_orig * mask + (1. - mask) * img
outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
quantize_denoised=quantize_denoised, temperature=temperature,
noise_dropout=noise_dropout, score_corrector=score_corrector,
corrector_kwargs=corrector_kwargs,
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=unconditional_conditioning,
old_eps=old_eps, t_next=ts_next)
img, pred_x0, e_t = outs
old_eps.append(e_t)
if len(old_eps) >= 4:
old_eps.pop(0)
if callback: callback(i)
if img_callback: img_callback(pred_x0, i)
return img
@torch.no_grad()
def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None):
b, *_, device = *x.shape, x.device
def get_model_output(x, t):
if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
e_t = self.apply_model(x, t, c)
else:
x_in = torch.cat([x] * 2)
t_in = torch.cat([t] * 2)
c_in = torch.cat([unconditional_conditioning, c])
e_t_uncond, e_t = self.apply_model(x_in, t_in, c_in).chunk(2)
e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
if score_corrector is not None:
assert self.parameterization == "eps"
e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
return e_t
alphas = self.alphas_cumprod if use_original_steps else self.ddim_alphas
alphas_prev = self.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
sqrt_one_minus_alphas = self.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
sigmas = self.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
def get_x_prev_and_pred_x0(e_t, index):
# select parameters corresponding to the currently considered timestep
a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
# current prediction for x_0
pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
if quantize_denoised:
pred_x0, _, *_ = self.first_stage_model.quantize(pred_x0)
# direction pointing to x_t
dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
if noise_dropout > 0.:
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
return x_prev, pred_x0
e_t = get_model_output(x, t)
if len(old_eps) == 0:
# Pseudo Improved Euler (2nd order)
x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
e_t_next = get_model_output(x_prev, t_next)
e_t_prime = (e_t + e_t_next) / 2
elif len(old_eps) == 1:
# 2nd order Pseudo Linear Multistep (Adams-Bashforth)
e_t_prime = (3 * e_t - old_eps[-1]) / 2
elif len(old_eps) == 2:
# 3nd order Pseudo Linear Multistep (Adams-Bashforth)
e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
elif len(old_eps) >= 3:
# 4nd order Pseudo Linear Multistep (Adams-Bashforth)
e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24
x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
return x_prev, pred_x0, e_t
@torch.no_grad()
def stochastic_encode(self, x0, t, seed, ddim_eta,ddim_steps,use_original_steps=False, noise=None, mask=None):
# fast, but does not allow for exact reconstruction
# t serves as an index to gather the correct alphas
self.make_schedule(ddim_num_steps=ddim_steps, ddim_eta=ddim_eta, verbose=False)
if use_original_steps:
sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
else:
sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
if noise is None:
b0, b1, b2, b3 = x0.shape
img_shape = (1, b1, b2, b3)
tens = []
print("seeds used = ", [seed+s for s in range(b0)])
for _ in range(b0):
torch.manual_seed(seed)
tens.append(torch.randn(img_shape, device=x0.device))
seed+=1
noise = torch.cat(tens)
del tens
if mask is not None:
noise = noise*mask
return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
extract_into_tensor(sqrt_one_minus_alphas_cumprod.to(self.cdevice), t, x0.shape) * noise)
@torch.no_grad()
def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
mask = None,use_original_steps=False):
if(self.turbo):
self.model1.to(self.cdevice)
self.model2.to(self.cdevice)
timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
timesteps = timesteps[:t_start]
time_range = np.flip(timesteps)
total_steps = timesteps.shape[0]
print(f"Running DDIM Sampling with {total_steps} timesteps")
iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
x_dec = x_latent
# x0 = x_latent
for i, step in enumerate(iterator):
index = total_steps - i - 1
ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
# if mask is not None:
# x_dec = x0 * mask + (1. - mask) * x_dec
x_dec = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=unconditional_conditioning)
# if mask is not None:
# return x0 * mask + (1. - mask) * x_dec
if(self.turbo):
self.model1.to("cpu")
self.model2.to("cpu")
return x_dec
@torch.no_grad()
def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
unconditional_guidance_scale=1., unconditional_conditioning=None):
b, *_, device = *x.shape, x.device
if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
e_t = self.apply_model(x, t, c)
else:
x_in = torch.cat([x] * 2)
t_in = torch.cat([t] * 2)
c_in = torch.cat([unconditional_conditioning, c])
# print("xin shape = ", x_in.shape)
e_t_uncond, e_t = self.apply_model(x_in, t_in, c_in).chunk(2)
e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
if score_corrector is not None:
assert self.model.parameterization == "eps"
e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
# select parameters corresponding to the currently considered timestep
a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
# current prediction for x_0
pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
if quantize_denoised:
pred_x0, _, *_ = self.first_stage_model.quantize(pred_x0)
# direction pointing to x_t
dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
if noise_dropout > 0.:
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
return x_prev

807
optimizedSD/openaimodelSplit.py Normal file
View File

@ -0,0 +1,807 @@
from abc import abstractmethod
import math
import numpy as np
import torch as th
import torch.nn as nn
import torch.nn.functional as F
from ldm.modules.diffusionmodules.util import (
checkpoint,
conv_nd,
linear,
avg_pool_nd,
zero_module,
normalization,
timestep_embedding,
)
from ldm.modules.attention import SpatialTransformer
class AttentionPool2d(nn.Module):
"""
Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
"""
def __init__(
self,
spacial_dim: int,
embed_dim: int,
num_heads_channels: int,
output_dim: int = None,
):
super().__init__()
self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5)
self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1)
self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1)
self.num_heads = embed_dim // num_heads_channels
self.attention = QKVAttention(self.num_heads)
def forward(self, x):
b, c, *_spatial = x.shape
x = x.reshape(b, c, -1) # NC(HW)
x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) # NC(HW+1)
x = x + self.positional_embedding[None, :, :].to(x.dtype) # NC(HW+1)
x = self.qkv_proj(x)
x = self.attention(x)
x = self.c_proj(x)
return x[:, :, 0]
class TimestepBlock(nn.Module):
"""
Any module where forward() takes timestep embeddings as a second argument.
"""
@abstractmethod
def forward(self, x, emb):
"""
Apply the module to `x` given `emb` timestep embeddings.
"""
class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
"""
A sequential module that passes timestep embeddings to the children that
support it as an extra input.
"""
def forward(self, x, emb, context=None):
for layer in self:
if isinstance(layer, TimestepBlock):
x = layer(x, emb)
elif isinstance(layer, SpatialTransformer):
x = layer(x, context)
else:
x = layer(x)
return x
class Upsample(nn.Module):
"""
An upsampling layer with an optional convolution.
:param channels: channels in the inputs and outputs.
:param use_conv: a bool determining if a convolution is applied.
:param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
upsampling occurs in the inner-two dimensions.
"""
def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
self.use_conv = use_conv
self.dims = dims
if use_conv:
self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
def forward(self, x):
assert x.shape[1] == self.channels
if self.dims == 3:
x = F.interpolate(
x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
)
else:
x = F.interpolate(x, scale_factor=2, mode="nearest")
if self.use_conv:
x = self.conv(x)
return x
class TransposedUpsample(nn.Module):
'Learned 2x upsampling without padding'
def __init__(self, channels, out_channels=None, ks=5):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2)
def forward(self,x):
return self.up(x)
class Downsample(nn.Module):
"""
A downsampling layer with an optional convolution.
:param channels: channels in the inputs and outputs.
:param use_conv: a bool determining if a convolution is applied.
:param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
downsampling occurs in the inner-two dimensions.
"""
def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
self.use_conv = use_conv
self.dims = dims
stride = 2 if dims != 3 else (1, 2, 2)
if use_conv:
self.op = conv_nd(
dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
)
else:
assert self.channels == self.out_channels
self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
def forward(self, x):
assert x.shape[1] == self.channels
return self.op(x)
class ResBlock(TimestepBlock):
"""
A residual block that can optionally change the number of channels.
:param channels: the number of input channels.
:param emb_channels: the number of timestep embedding channels.
:param dropout: the rate of dropout.
:param out_channels: if specified, the number of out channels.
:param use_conv: if True and out_channels is specified, use a spatial
convolution instead of a smaller 1x1 convolution to change the
channels in the skip connection.
:param dims: determines if the signal is 1D, 2D, or 3D.
:param use_checkpoint: if True, use gradient checkpointing on this module.
:param up: if True, use this block for upsampling.
:param down: if True, use this block for downsampling.
"""
def __init__(
self,
channels,
emb_channels,
dropout,
out_channels=None,
use_conv=False,
use_scale_shift_norm=False,
dims=2,
use_checkpoint=False,
up=False,
down=False,
):
super().__init__()
self.channels = channels
self.emb_channels = emb_channels
self.dropout = dropout
self.out_channels = out_channels or channels
self.use_conv = use_conv
self.use_checkpoint = use_checkpoint
self.use_scale_shift_norm = use_scale_shift_norm
self.in_layers = nn.Sequential(
normalization(channels),
nn.SiLU(),
conv_nd(dims, channels, self.out_channels, 3, padding=1),
)
self.updown = up or down
if up:
self.h_upd = Upsample(channels, False, dims)
self.x_upd = Upsample(channels, False, dims)
elif down:
self.h_upd = Downsample(channels, False, dims)
self.x_upd = Downsample(channels, False, dims)
else:
self.h_upd = self.x_upd = nn.Identity()
self.emb_layers = nn.Sequential(
nn.SiLU(),
linear(
emb_channels,
2 * self.out_channels if use_scale_shift_norm else self.out_channels,
),
)
self.out_layers = nn.Sequential(
normalization(self.out_channels),
nn.SiLU(),
nn.Dropout(p=dropout),
zero_module(
conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
),
)
if self.out_channels == channels:
self.skip_connection = nn.Identity()
elif use_conv:
self.skip_connection = conv_nd(
dims, channels, self.out_channels, 3, padding=1
)
else:
self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
def forward(self, x, emb):
"""
Apply the block to a Tensor, conditioned on a timestep embedding.
:param x: an [N x C x ...] Tensor of features.
:param emb: an [N x emb_channels] Tensor of timestep embeddings.
:return: an [N x C x ...] Tensor of outputs.
"""
return checkpoint(
self._forward, (x, emb), self.parameters(), self.use_checkpoint
)
def _forward(self, x, emb):
if self.updown:
in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
h = in_rest(x)
h = self.h_upd(h)
x = self.x_upd(x)
h = in_conv(h)
else:
h = self.in_layers(x)
emb_out = self.emb_layers(emb).type(h.dtype)
while len(emb_out.shape) < len(h.shape):
emb_out = emb_out[..., None]
if self.use_scale_shift_norm:
out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
scale, shift = th.chunk(emb_out, 2, dim=1)
h = out_norm(h) * (1 + scale) + shift
h = out_rest(h)
else:
h = h + emb_out
h = self.out_layers(h)
return self.skip_connection(x) + h
class AttentionBlock(nn.Module):
"""
An attention block that allows spatial positions to attend to each other.
Originally ported from here, but adapted to the N-d case.
https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
"""
def __init__(
self,
channels,
num_heads=1,
num_head_channels=-1,
use_checkpoint=False,
use_new_attention_order=False,
):
super().__init__()
self.channels = channels
if num_head_channels == -1:
self.num_heads = num_heads
else:
assert (
channels % num_head_channels == 0
), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
self.num_heads = channels // num_head_channels
self.use_checkpoint = use_checkpoint
self.norm = normalization(channels)
self.qkv = conv_nd(1, channels, channels * 3, 1)
if use_new_attention_order:
# split qkv before split heads
self.attention = QKVAttention(self.num_heads)
else:
# split heads before split qkv
self.attention = QKVAttentionLegacy(self.num_heads)
self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
def forward(self, x):
return checkpoint(self._forward, (x,), self.parameters(), True) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!!
#return pt_checkpoint(self._forward, x) # pytorch
def _forward(self, x):
b, c, *spatial = x.shape
x = x.reshape(b, c, -1)
qkv = self.qkv(self.norm(x))
h = self.attention(qkv)
h = self.proj_out(h)
return (x + h).reshape(b, c, *spatial)
def count_flops_attn(model, _x, y):
"""
A counter for the `thop` package to count the operations in an
attention operation.
Meant to be used like:
macs, params = thop.profile(
model,
inputs=(inputs, timestamps),
custom_ops={QKVAttention: QKVAttention.count_flops},
)
"""
b, c, *spatial = y[0].shape
num_spatial = int(np.prod(spatial))
# We perform two matmuls with the same number of ops.
# The first computes the weight matrix, the second computes
# the combination of the value vectors.
matmul_ops = 2 * b * (num_spatial ** 2) * c
model.total_ops += th.DoubleTensor([matmul_ops])
class QKVAttentionLegacy(nn.Module):
"""
A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
"""
def __init__(self, n_heads):
super().__init__()
self.n_heads = n_heads
def forward(self, qkv):
"""
Apply QKV attention.
:param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
:return: an [N x (H * C) x T] tensor after attention.
"""
bs, width, length = qkv.shape
assert width % (3 * self.n_heads) == 0
ch = width // (3 * self.n_heads)
q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
scale = 1 / math.sqrt(math.sqrt(ch))
weight = th.einsum(
"bct,bcs->bts", q * scale, k * scale
) # More stable with f16 than dividing afterwards
weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
a = th.einsum("bts,bcs->bct", weight, v)
return a.reshape(bs, -1, length)
@staticmethod
def count_flops(model, _x, y):
return count_flops_attn(model, _x, y)
class QKVAttention(nn.Module):
"""
A module which performs QKV attention and splits in a different order.
"""
def __init__(self, n_heads):
super().__init__()
self.n_heads = n_heads
def forward(self, qkv):
"""
Apply QKV attention.
:param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
:return: an [N x (H * C) x T] tensor after attention.
"""
bs, width, length = qkv.shape
assert width % (3 * self.n_heads) == 0
ch = width // (3 * self.n_heads)
q, k, v = qkv.chunk(3, dim=1)
scale = 1 / math.sqrt(math.sqrt(ch))
weight = th.einsum(
"bct,bcs->bts",
(q * scale).view(bs * self.n_heads, ch, length),
(k * scale).view(bs * self.n_heads, ch, length),
) # More stable with f16 than dividing afterwards
weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length))
return a.reshape(bs, -1, length)
@staticmethod
def count_flops(model, _x, y):
return count_flops_attn(model, _x, y)
class UNetModelEncode(nn.Module):
def __init__(
self,
image_size,
in_channels,
model_channels,
out_channels,
num_res_blocks,
attention_resolutions,
dropout=0,
channel_mult=(1, 2, 4, 8),
conv_resample=True,
dims=2,
num_classes=None,
use_checkpoint=False,
use_fp16=False,
num_heads=-1,
num_head_channels=-1,
num_heads_upsample=-1,
use_scale_shift_norm=False,
resblock_updown=False,
use_new_attention_order=False,
use_spatial_transformer=False, # custom transformer support
transformer_depth=1, # custom transformer support
context_dim=None, # custom transformer support
n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model
legacy=True,
):
super().__init__()
if use_spatial_transformer:
assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
if context_dim is not None:
assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
from omegaconf.listconfig import ListConfig
if type(context_dim) == ListConfig:
context_dim = list(context_dim)
if num_heads_upsample == -1:
num_heads_upsample = num_heads
if num_heads == -1:
assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
if num_head_channels == -1:
assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
self.image_size = image_size
self.in_channels = in_channels
self.model_channels = model_channels
self.out_channels = out_channels
self.num_res_blocks = num_res_blocks
self.attention_resolutions = attention_resolutions
self.dropout = dropout
self.channel_mult = channel_mult
self.conv_resample = conv_resample
self.num_classes = num_classes
self.use_checkpoint = use_checkpoint
self.dtype = th.float16 if use_fp16 else th.float32
self.num_heads = num_heads
self.num_head_channels = num_head_channels
self.num_heads_upsample = num_heads_upsample
self.predict_codebook_ids = n_embed is not None
time_embed_dim = model_channels * 4
self.time_embed = nn.Sequential(
linear(model_channels, time_embed_dim),
nn.SiLU(),
linear(time_embed_dim, time_embed_dim),
)
if self.num_classes is not None:
self.label_emb = nn.Embedding(num_classes, time_embed_dim)
self.input_blocks = nn.ModuleList(
[
TimestepEmbedSequential(
conv_nd(dims, in_channels, model_channels, 3, padding=1)
)
]
)
self._feature_size = model_channels
input_block_chans = [model_channels]
ch = model_channels
ds = 1
for level, mult in enumerate(channel_mult):
for _ in range(num_res_blocks):
layers = [
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=mult * model_channels,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
)
]
ch = mult * model_channels
if ds in attention_resolutions:
if num_head_channels == -1:
dim_head = ch // num_heads
else:
num_heads = ch // num_head_channels
dim_head = num_head_channels
if legacy:
#num_heads = 1
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
layers.append(
AttentionBlock(
ch,
use_checkpoint=use_checkpoint,
num_heads=num_heads,
num_head_channels=dim_head,
use_new_attention_order=use_new_attention_order,
) if not use_spatial_transformer else SpatialTransformer(
ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
)
)
self.input_blocks.append(TimestepEmbedSequential(*layers))
self._feature_size += ch
input_block_chans.append(ch)
if level != len(channel_mult) - 1:
out_ch = ch
self.input_blocks.append(
TimestepEmbedSequential(
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=out_ch,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
down=True,
)
if resblock_updown
else Downsample(
ch, conv_resample, dims=dims, out_channels=out_ch
)
)
)
ch = out_ch
input_block_chans.append(ch)
ds *= 2
self._feature_size += ch
if num_head_channels == -1:
dim_head = ch // num_heads
else:
num_heads = ch // num_head_channels
dim_head = num_head_channels
if legacy:
#num_heads = 1
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
self.middle_block = TimestepEmbedSequential(
ResBlock(
ch,
time_embed_dim,
dropout,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
),
AttentionBlock(
ch,
use_checkpoint=use_checkpoint,
num_heads=num_heads,
num_head_channels=dim_head,
use_new_attention_order=use_new_attention_order,
) if not use_spatial_transformer else SpatialTransformer(
ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
),
ResBlock(
ch,
time_embed_dim,
dropout,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
),
)
self._feature_size += ch
def forward(self, x, timesteps=None, context=None, y=None):
"""
Apply the model to an input batch.
:param x: an [N x C x ...] Tensor of inputs.
:param timesteps: a 1-D batch of timesteps.
:param context: conditioning plugged in via crossattn
:param y: an [N] Tensor of labels, if class-conditional.
:return: an [N x C x ...] Tensor of outputs.
"""
assert (y is not None) == (
self.num_classes is not None
), "must specify y if and only if the model is class-conditional"
hs = []
t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
emb = self.time_embed(t_emb)
if self.num_classes is not None:
assert y.shape == (x.shape[0],)
emb = emb + self.label_emb(y)
h = x.type(self.dtype)
for module in self.input_blocks:
h = module(h, emb, context)
hs.append(h)
h = self.middle_block(h, emb, context)
return h, emb, hs
class UNetModelDecode(nn.Module):
def __init__(
self,
image_size,
in_channels,
model_channels,
out_channels,
num_res_blocks,
attention_resolutions,
dropout=0,
channel_mult=(1, 2, 4, 8),
conv_resample=True,
dims=2,
num_classes=None,
use_checkpoint=False,
use_fp16=False,
num_heads=-1,
num_head_channels=-1,
num_heads_upsample=-1,
use_scale_shift_norm=False,
resblock_updown=False,
use_new_attention_order=False,
use_spatial_transformer=False, # custom transformer support
transformer_depth=1, # custom transformer support
context_dim=None, # custom transformer support
n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model
legacy=True,
):
super().__init__()
if use_spatial_transformer:
assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
if context_dim is not None:
assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
from omegaconf.listconfig import ListConfig
if type(context_dim) == ListConfig:
context_dim = list(context_dim)
if num_heads_upsample == -1:
num_heads_upsample = num_heads
if num_heads == -1:
assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
if num_head_channels == -1:
assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
self.image_size = image_size
self.in_channels = in_channels
self.model_channels = model_channels
self.out_channels = out_channels
self.num_res_blocks = num_res_blocks
self.attention_resolutions = attention_resolutions
self.dropout = dropout
self.channel_mult = channel_mult
self.conv_resample = conv_resample
self.num_classes = num_classes
self.use_checkpoint = use_checkpoint
self.dtype = th.float16 if use_fp16 else th.float32
self.num_heads = num_heads
self.num_head_channels = num_head_channels
self.num_heads_upsample = num_heads_upsample
self.predict_codebook_ids = n_embed is not None
time_embed_dim = model_channels * 4
self._feature_size = model_channels
input_block_chans = [model_channels]
ch = model_channels
ds = 1
for level, mult in enumerate(channel_mult):
for _ in range(num_res_blocks):
ch = mult * model_channels
if ds in attention_resolutions:
if num_head_channels == -1:
dim_head = ch // num_heads
else:
num_heads = ch // num_head_channels
dim_head = num_head_channels
if legacy:
#num_heads = 1
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
self._feature_size += ch
input_block_chans.append(ch)
if level != len(channel_mult) - 1:
out_ch = ch
ch = out_ch
input_block_chans.append(ch)
ds *= 2
self._feature_size += ch
if num_head_channels == -1:
dim_head = ch // num_heads
else:
num_heads = ch // num_head_channels
dim_head = num_head_channels
if legacy:
#num_heads = 1
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
self._feature_size += ch
self.output_blocks = nn.ModuleList([])
for level, mult in list(enumerate(channel_mult))[::-1]:
for i in range(num_res_blocks + 1):
ich = input_block_chans.pop()
layers = [
ResBlock(
ch + ich,
time_embed_dim,
dropout,
out_channels=model_channels * mult,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
)
]
ch = model_channels * mult
if ds in attention_resolutions:
if num_head_channels == -1:
dim_head = ch // num_heads
else:
num_heads = ch // num_head_channels
dim_head = num_head_channels
if legacy:
#num_heads = 1
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
layers.append(
AttentionBlock(
ch,
use_checkpoint=use_checkpoint,
num_heads=num_heads_upsample,
num_head_channels=dim_head,
use_new_attention_order=use_new_attention_order,
) if not use_spatial_transformer else SpatialTransformer(
ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
)
)
if level and i == num_res_blocks:
out_ch = ch
layers.append(
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=out_ch,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
up=True,
)
if resblock_updown
else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
)
ds //= 2
self.output_blocks.append(TimestepEmbedSequential(*layers))
self._feature_size += ch
self.out = nn.Sequential(
normalization(ch),
nn.SiLU(),
zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
)
if self.predict_codebook_ids:
self.id_predictor = nn.Sequential(
normalization(ch),
conv_nd(dims, model_channels, n_embed, 1),
#nn.LogSoftmax(dim=1) # change to cross_entropy and produce non-normalized logits
)
def forward(self, h,emb,tp,hs, context=None, y=None):
"""
Apply the model to an input batch.
:param x: an [N x C x ...] Tensor of inputs.
:param timesteps: a 1-D batch of timesteps.
:param context: conditioning plugged in via crossattn
:param y: an [N] Tensor of labels, if class-conditional.
:return: an [N x C x ...] Tensor of outputs.
"""
for module in self.output_blocks:
h = th.cat([h, hs.pop()], dim=1)
h = module(h, emb, context)
h = h.type(tp)
if self.predict_codebook_ids:
return self.id_predictor(h)
else:
return self.out(h)

73
optimizedSD/optimUtils.py Normal file
View File

@ -0,0 +1,73 @@
import os
import pandas as pd
def split_weighted_subprompts(text):
"""
grabs all text up to the first occurrence of ':'
uses the grabbed text as a sub-prompt, and takes the value following ':' as weight
if ':' has no value defined, defaults to 1.0
repeats until no text remaining
"""
remaining = len(text)
prompts = []
weights = []
while remaining > 0:
if ":" in text:
idx = text.index(":") # first occurrence from start
# grab up to index as sub-prompt
prompt = text[:idx]
remaining -= idx
# remove from main text
text = text[idx+1:]
# find value for weight
if " " in text:
idx = text.index(" ") # first occurence
else: # no space, read to end
idx = len(text)
if idx != 0:
try:
weight = float(text[:idx])
except: # couldn't treat as float
print(f"Warning: '{text[:idx]}' is not a value, are you missing a space?")
weight = 1.0
else: # no value found
weight = 1.0
# remove from main text
remaining -= idx
text = text[idx+1:]
# append the sub-prompt and its weight
prompts.append(prompt)
weights.append(weight)
else: # no : found
if len(text) > 0: # there is still text though
# take remainder as weight 1
prompts.append(text)
weights.append(1.0)
remaining = 0
return prompts, weights
def logger(params, log_csv):
os.makedirs('logs', exist_ok=True)
cols = [arg for arg, _ in params.items()]
if not os.path.exists(log_csv):
df = pd.DataFrame(columns=cols)
df.to_csv(log_csv, index=False)
df = pd.read_csv(log_csv)
for arg in cols:
if arg not in df.columns:
df[arg] = ""
df.to_csv(log_csv, index = False)
li = {}
cols = [col for col in df.columns]
data = {arg:value for arg, value in params.items()}
for col in cols:
if col in data:
li[col] = data[col]
else:
li[col] = ''
df = pd.DataFrame(li,index = [0])
df.to_csv(log_csv,index=False, mode='a', header=False)

297
optimizedSD/optimized_txt2img.py Normal file
View File

@ -0,0 +1,297 @@
import argparse, os, sys, glob, random
import torch
import numpy as np
import copy
from omegaconf import OmegaConf
from PIL import Image
from tqdm import tqdm, trange
from itertools import islice
from einops import rearrange
from torchvision.utils import make_grid
import time
from pytorch_lightning import seed_everything
from torch import autocast
from contextlib import contextmanager, nullcontext
from ldm.util import instantiate_from_config
def chunk(it, size):
it = iter(it)
return iter(lambda: tuple(islice(it, size)), ())
def load_model_from_config(ckpt, verbose=False):
print(f"Loading model from {ckpt}")
pl_sd = torch.load(ckpt, map_location="cpu")
if "global_step" in pl_sd:
print(f"Global Step: {pl_sd['global_step']}")
sd = pl_sd["state_dict"]
return sd
config = "optimizedSD/v1-inference.yaml"
ckpt = "models/ldm/stable-diffusion-v1/model.ckpt"
device = "cuda"
parser = argparse.ArgumentParser()
parser.add_argument(
"--prompt",
type=str,
nargs="?",
default="a painting of a virus monster playing guitar",
help="the prompt to render"
)
parser.add_argument(
"--outdir",
type=str,
nargs="?",
help="dir to write results to",
default="outputs/txt2img-samples"
)
parser.add_argument(
"--skip_grid",
action='store_true',
help="do not save a grid, only individual samples. Helpful when evaluating lots of samples",
)
parser.add_argument(
"--skip_save",
action='store_true',
help="do not save individual samples. For speed measurements.",
)
parser.add_argument(
"--ddim_steps",
type=int,
default=50,
help="number of ddim sampling steps",
)
parser.add_argument(
"--fixed_code",
action='store_true',
help="if enabled, uses the same starting code across samples ",
)
parser.add_argument(
"--ddim_eta",
type=float,
default=0.0,
help="ddim eta (eta=0.0 corresponds to deterministic sampling",
)
parser.add_argument(
"--n_iter",
type=int,
default=1,
help="sample this often",
)
parser.add_argument(
"--H",
type=int,
default=512,
help="image height, in pixel space",
)
parser.add_argument(
"--W",
type=int,
default=512,
help="image width, in pixel space",
)
parser.add_argument(
"--C",
type=int,
default=4,
help="latent channels",
)
parser.add_argument(
"--f",
type=int,
default=8,
help="downsampling factor",
)
parser.add_argument(
"--n_samples",
type=int,
default=5,
help="how many samples to produce for each given prompt. A.k.a. batch size",
)
parser.add_argument(
"--n_rows",
type=int,
default=0,
help="rows in the grid (default: n_samples)",
)
parser.add_argument(
"--scale",
type=float,
default=7.5,
help="unconditional guidance scale: eps = eps(x, empty) + scale * (eps(x, cond) - eps(x, empty))",
)
parser.add_argument(
"--from-file",
type=str,
help="if specified, load prompts from this file",
)
parser.add_argument(
"--seed",
type=int,
default=42,
help="the seed (for reproducible sampling)",
)
parser.add_argument(
"--small_batch",
action='store_true',
help="Reduce inference time when generate a smaller batch of images",
)
parser.add_argument(
"--precision",
type=str,
help="evaluate at this precision",
choices=["full", "autocast"],
default="autocast"
)
opt = parser.parse_args()
tic = time.time()
os.makedirs(opt.outdir, exist_ok=True)
outpath = opt.outdir
sample_path = os.path.join(outpath, "samples", "_".join(opt.prompt.split())[:255])
os.makedirs(sample_path, exist_ok=True)
base_count = len(os.listdir(sample_path))
grid_count = len(os.listdir(outpath)) - 1
seed_everything(opt.seed)
sd = load_model_from_config(f"{ckpt}")
li = []
lo = []
for key, value in sd.items():
sp = key.split('.')
if(sp[0]) == 'model':
if('input_blocks' in sp):
li.append(key)
elif('middle_block' in sp):
li.append(key)
elif('time_embed' in sp):
li.append(key)
else:
lo.append(key)
for key in li:
sd['model1.' + key[6:]] = sd.pop(key)
for key in lo:
sd['model2.' + key[6:]] = sd.pop(key)
config = OmegaConf.load(f"{config}")
config.modelUNet.params.ddim_steps = opt.ddim_steps
if opt.small_batch:
config.modelUNet.params.small_batch = True
else:
config.modelUNet.params.small_batch = False
model = instantiate_from_config(config.modelUNet)
_, _ = model.load_state_dict(sd, strict=False)
model.eval()
modelCS = instantiate_from_config(config.modelCondStage)
_, _ = modelCS.load_state_dict(sd, strict=False)
modelCS.eval()
modelFS = instantiate_from_config(config.modelFirstStage)
_, _ = modelFS.load_state_dict(sd, strict=False)
modelFS.eval()
if opt.precision == "autocast":
model.half()
modelCS.half()
start_code = None
if opt.fixed_code:
start_code = torch.randn([opt.n_samples, opt.C, opt.H // opt.f, opt.W // opt.f], device=device)
batch_size = opt.n_samples
n_rows = opt.n_rows if opt.n_rows > 0 else batch_size
if not opt.from_file:
prompt = opt.prompt
assert prompt is not None
data = [batch_size * [prompt]]
else:
print(f"reading prompts from {opt.from_file}")
with open(opt.from_file, "r") as f:
data = f.read().splitlines()
data = list(chunk(data, batch_size))
precision_scope = autocast if opt.precision=="autocast" else nullcontext
with torch.no_grad():
all_samples = list()
for n in trange(opt.n_iter, desc="Sampling"):
for prompts in tqdm(data, desc="data"):
with precision_scope("cuda"):
modelCS.to(device)
uc = None
if opt.scale != 1.0:
uc = modelCS.get_learned_conditioning(batch_size * [""])
if isinstance(prompts, tuple):
prompts = list(prompts)
c = modelCS.get_learned_conditioning(prompts)
shape = [opt.C, opt.H // opt.f, opt.W // opt.f]
mem = torch.cuda.memory_allocated()/1e6
modelCS.to("cpu")
while(torch.cuda.memory_allocated()/1e6 >= mem):
time.sleep(1)
samples_ddim = model.sample(S=opt.ddim_steps,
conditioning=c,
batch_size=opt.n_samples,
shape=shape,
verbose=False,
unconditional_guidance_scale=opt.scale,
unconditional_conditioning=uc,
eta=opt.ddim_eta,
x_T=start_code)
modelFS.to(device)
print("saving images")
for i in range(batch_size):
x_samples_ddim = modelFS.decode_first_stage(samples_ddim[i].unsqueeze(0))
x_sample = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
# for x_sample in x_samples_ddim:
x_sample = 255. * rearrange(x_sample[0].cpu().numpy(), 'c h w -> h w c')
Image.fromarray(x_sample.astype(np.uint8)).save(
os.path.join(sample_path, f"{base_count:05}.png"))
base_count += 1
mem = torch.cuda.memory_allocated()/1e6
modelFS.to("cpu")
while(torch.cuda.memory_allocated()/1e6 >= mem):
time.sleep(1)
# if not opt.skip_grid:
# all_samples.append(x_samples_ddim)
del samples_ddim
print("memory_final = ", torch.cuda.memory_allocated()/1e6)
# if not skip_grid:
# # additionally, save as grid
# grid = torch.stack(all_samples, 0)
# grid = rearrange(grid, 'n b c h w -> (n b) c h w')
# grid = make_grid(grid, nrow=n_rows)
# # to image
# grid = 255. * rearrange(grid, 'c h w -> h w c').cpu().numpy()
# Image.fromarray(grid.astype(np.uint8)).save(os.path.join(outpath, f'grid-{grid_count:04}.png'))
# grid_count += 1
toc = time.time()
time_taken = (toc-tic)/60.0
print(("Your samples are ready in {0:.2f} minutes and waiting for you here \n" + sample_path).format(time_taken))

114
optimizedSD/v1-inference.yaml Normal file
View File

@ -0,0 +1,114 @@
modelUNet:
base_learning_rate: 1.0e-04
target: optimizedSD.ddpm.UNet
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
image_size: 64
channels: 4
cond_stage_trainable: false # Note: different from the one we trained before
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
unetConfigEncode:
target: optimizedSD.openaimodelSplit.UNetModelEncode
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [4, 2, 1]
num_res_blocks: 2
channel_mult: [1, 2, 4, 4]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unetConfigDecode:
target: optimizedSD.openaimodelSplit.UNetModelDecode
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [4, 2, 1]
num_res_blocks: 2
channel_mult: [1, 2, 4, 4]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
modelFirstStage:
target: optimizedSD.ddpm.FirstStage
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
image_size: 64
channels: 4
cond_stage_trainable: false # Note: different from the one we trained before
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
modelCondStage:
target: optimizedSD.ddpm.CondStage
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
image_size: 64
channels: 4
cond_stage_trainable: false # Note: different from the one we trained before
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
params:
device: cpu

BIN
screenshot.png
View File

Binary file not shown.
Side by Side

Before

Width:  |  Height:  |  Size: 865 KiB

128
relauncher.py → scripts/relauncher.py
View File

@ -1,64 +1,64 @@
import os, time
# USER CHANGABLE ARGUMENTS
# Change to `True` if you wish to enable these common arguments
# Run upscaling models on the CPU
extra_models_cpu = False
# Automatically open a new browser window or tab on first launch
open_in_browser = False
# Run Stable Diffusion in Optimized Mode - Only requires 4Gb of VRAM, but is significantly slower
optimized = False
# Run in Optimized Turbo Mode - Needs more VRAM than regular optimized mode, but is faster
optimized_turbo = False
# Creates a public xxxxx.gradio.app share link to allow others to use your interface (requires properly forwarded ports to work correctly)
share = False
# Enter other `--arguments` you wish to use - Must be entered as a `--argument ` syntax
additional_arguments = ""
# BEGIN RELAUNCHER PYTHON CODE
common_arguments = ""
if extra_models_cpu == True:
common_arguments += "--extra-models-cpu "
if optimized_turbo == True:
common_arguments += "--optimized-turbo "
if optimized == True:
common_arguments += "--optimized "
if share == True:
common_arguments += "--share "
if open_in_browser == True:
inbrowser_argument = "--inbrowser "
else:
inbrowser_argument = ""
n = 0
while True:
if n == 0:
print('Relauncher: Launching...')
os.system(f"python scripts/webui.py {common_arguments} {inbrowser_argument} {additional_arguments}")
else:
print(f'\tRelaunch count: {n}')
print('Relauncher: Launching...')
os.system(f"python scripts/webui.py {common_arguments} {additional_arguments}")
n += 1
if n > 100:
print ('Too many relaunch attempts. Aborting...')
break
print('Relauncher: Process is ending. Relaunching in 1s...')
time.sleep(1)
import os, time
# USER CHANGABLE ARGUMENTS
# Change to `True` if you wish to enable these common arguments
# Run upscaling models on the CPU
extra_models_cpu = False
# Automatically open a new browser window or tab on first launch
open_in_browser = False
# Run Stable Diffusion in Optimized Mode - Only requires 4Gb of VRAM, but is significantly slower
optimized = False
# Run in Optimized Turbo Mode - Needs more VRAM than regular optimized mode, but is faster
optimized_turbo = False
# Creates a public xxxxx.gradio.app share link to allow others to use your interface (requires properly forwarded ports to work correctly)
share = False
# Enter other `--arguments` you wish to use - Must be entered as a `--argument ` syntax
additional_arguments = ""
# BEGIN RELAUNCHER PYTHON CODE
common_arguments = ""
if extra_models_cpu == True:
common_arguments += "--extra-models-cpu "
if optimized_turbo == True:
common_arguments += "--optimized-turbo "
if optimized == True:
common_arguments += "--optimized "
if share == True:
common_arguments += "--share "
if open_in_browser == True:
inbrowser_argument = "--inbrowser "
else:
inbrowser_argument = ""
n = 0
while True:
if n == 0:
print('Relauncher: Launching...')
os.system(f"python scripts/webui.py {common_arguments} {inbrowser_argument} {additional_arguments}")
else:
print(f'\tRelaunch count: {n}')
print('Relauncher: Launching...')
os.system(f"python scripts/webui.py {common_arguments} {additional_arguments}")
n += 1
if n > 100:
print ('Too many relaunch attempts. Aborting...')
break
print('Relauncher: Process is ending. Relaunching in 1s...')
time.sleep(1)

4432
webui.py → scripts/webui.py
View File

File diff suppressed because it is too large Load Diff

13
setup.py Normal file
View File

@ -0,0 +1,13 @@
from setuptools import setup, find_packages
setup(
name='latent-diffusion',
version='0.0.1',
description='',
packages=find_packages(),
install_requires=[
'torch',
'numpy',
'tqdm',
],
)

64
sync_local.py
View File

@ -1,64 +0,0 @@
#!/usr/bin/env python
# sync_local.py
import argparse, os, sys, glob, re, yaml, shutil
parser = argparse.ArgumentParser()
parser.add_argument("-d", "--dest", type=str, help="dir to write files to. Required.", default=None)
parser.add_argument("-s", "--source", type=str, help="dir to read files from.", default=".")
parser.add_argument("-c", "--config", type=str, help="path to sync.yaml", default="./.github/sync.yml")
parser.add_argument("-k", "--key", type=str, help="key to read from sync.yaml. (Default: hlky/stable-diffusion)", default="hlky/stable-diffusion")
parser.add_argument("-r", "--reverse", action='store_true', help="Swaps source and destination to do a reverse copy.")
opt = parser.parse_args()
#Validate required parameters
if opt.dest is None:
raise SystemExit("Error: Missing --dest argument. Need to know where to copy files to.")
#Read config file
config_file = None
if opt.config is not None and os.path.isfile(opt.config):
try:
print(f"Reading copy list from {opt.config}")
with open(opt.config, "r", encoding="utf8") as f:
config_file = yaml.safe_load(f)
except (OSError, yaml.YAMLError) as e:
print(f"Error loading config file {opt.config}:", e, file=sys.stderr)
else:
raise SystemExit(f"Error: config file not found. Config value: {opt.config}")
#Double check config just in case
if config_file is None:
raise SystemExit("Unknown error why config_file not set")
#Pull key from config file
try:
print(f"Using key: {opt.key}")
copyList = config_file[opt.key]
except(KeyError) as e:
raise SystemExit(f"KeyError: could not find repo in config file. Looking for: {opt.key}")
#Copy each item pair in config file.
for i, copyItem in enumerate(copyList):
if(copyItem["dest"]) is None:
print(f"skip copy item {i}, missing dest")
continue
if(copyItem["source"]) is None:
print("skip copy item {i}, missing source")
continue
fullDest = os.path.join(opt.dest, copyItem["dest"])
fullSource = os.path.join(opt.source, copyItem["source"])
if opt.reverse:
fullDest, fullSource = fullSource, fullDest
try:
print(f"Copying file from source: [{fullSource}] to dest: [{fullDest}]")
os.makedirs(os.path.dirname(fullDest), exist_ok=True)
shutil.copy2(fullSource, fullDest)
except (OSError, IOError) as e:
print(f"Error copying file.", e, file=sys.stderr)
print("Done copying files.")

32
txt2img.yaml
View File

@ -1,32 +0,0 @@
# Example configuration file for CLI usage.
# Intended for advanced users.
target: txt2img # img2img not implemented.
# prompt can either be string or list.
# If prompt is list then the prompts are iterated in order.
prompt:
- E=mc2 # prompt 1
- hodl # prompt 2
ddim_steps: 50 # legacy name, applies to all algorithms.
# Adding an int to toggles enables the corresponding feature.
# 0: Create prompt matrix (separate multiple prompts using |, and get all combinations of them)
# 1: Normalize Prompt Weights (ensure sum of weights add up to 1.0)
# 2: Save individual images
# 3: Save grid
# 4: Sort samples by prompt
# 5: Write sample info files
# 6: Fix faces using GFPGAN
# 7: Upscale images using Real-ESRGAN (NOT TESTED)
toggles: [1, 2, 3, 4, 5]
sampler_name: k_lms # Valid: DDIM, k_dpm_2_a, k_dpm_2, k_euler_a, k_euler, k_heun, k_lms
ddim_eta: 0.0
n_iter: 1
batch_size: 1
cfg_scale: 7.5
seed: 3172010360 # Leave blank for random seed.
height: 512
width: 512
# The options below this line have not been tested. Use at your own discretion.
# Textual inversion embeddings file path:
fp:
realesrgan_model_name:

51
webui.cmd Normal file
View File

@ -0,0 +1,51 @@
@echo off
set conda_env_name=ldm
:: Put the path to conda directory after "=" sign if it's installed at non-standard path:
set custom_conda_path=
IF NOT "%custom_conda_path%"=="" (
set paths=%custom_conda_path%;%paths%
)
:: Put the path to conda directory in a file called "custom-conda-path.txt" if it's installed at non-standard path:
FOR /F %%i IN (custom-conda-path.txt) DO set custom_conda_path=%%i
set paths=%ProgramData%\miniconda3
set paths=%paths%;%USERPROFILE%\miniconda3
set paths=%paths%;%ProgramData%\anaconda3
set paths=%paths%;%USERPROFILE%\anaconda3
for %%a in (%paths%) do (
IF NOT "%custom_conda_path%"=="" (
set paths=%custom_conda_path%;%paths%
)
)
for %%a in (%paths%) do (
if EXIST "%%a\Scripts\activate.bat" (
SET CONDA_PATH=%%a
echo anaconda3/miniconda3 detected in %%a
goto :foundPath
)
)
IF "%CONDA_PATH%"=="" (
echo anaconda3/miniconda3 not found. Install from here https://docs.conda.io/en/latest/miniconda.html
exit /b 1
)
:foundPath
call "%CONDA_PATH%\Scripts\activate.bat"
call conda env create -n "%conda_env_name%" -f environment.yaml
call conda env update -n "%conda_env_name%" --file environment.yaml --prune
call "%CONDA_PATH%\Scripts\activate.bat" "%conda_env_name%"
python "%CD%"\scripts\relauncher.py
:PROMPT
set SETUPTOOLS_USE_DISTUTILS=stdlib
IF EXIST "models\ldm\stable-diffusion-v1\model.ckpt" (
python scripts/relauncher.py
) ELSE (
ECHO Your model file does not exist! Place it in 'models\ldm\stable-diffusion-v1' with the name 'model.ckpt'.
)

264
webui_playground.py
View File

@ -1,264 +0,0 @@
import gradio as gr
import time
import os
from PIL import Image, ImageFont, ImageDraw, ImageFilter, ImageOps
from io import BytesIO
import base64
import re
from frontend.frontend import draw_gradio_ui
from frontend.ui_functions import resize_image
"""
This file is here to play around with the interface without loading the whole model
TBD - extract all the UI into this file and import from the main webui.
"""
GFPGAN = True
RealESRGAN = True
def run_goBIG():
pass
def mock_processing(prompt: str, seed: str, width: int, height: int, steps: int,
cfg_scale: float, sampler: str, batch_count: int):
args_and_names = {
"seed": seed,
"width": width,
"height": height,
"steps": steps,
"cfg_scale": str(cfg_scale),
"sampler": sampler,
}
full_string = f"{prompt}\n" + " ".join([f"{k}:" for k, v in args_and_names.items()])
info = {
'text': full_string,
'entities': [
{'entity': str(v), 'start': full_string.find(f"{k}:"), 'end': full_string.find(f"{k}:") + len(f"{k} ")} for
k, v in args_and_names.items()]
}
images = []
for i in range(batch_count):
images.append(f"http://placeimg.com/{width}/{height}/any")
return images, int(time.time()), info, 'random output'
def txt2img(*args, **kwargs):
# Output should match output_txt2img_gallery, output_txt2img_seed, output_txt2img_params, output_txt2img_stats
# info = f"""{args[0]} --seed {args[9]} --W {args[11]} --H {args[10]} -s {args[1]} -C {float(args[8])} --sampler {args[2]} """.strip()
return mock_processing(
prompt=args[0],
seed=args[9],
width=args[11],
height=args[10],
steps=args[1],
cfg_scale=args[8],
sampler=args[2],
batch_count=args[6]
)
def img2img(*args, **kwargs):
return mock_processing(
prompt=args[0],
seed=args[12],
width=args[14],
height=args[13],
steps=args[5],
cfg_scale=args[10],
sampler=args[6],
batch_count=args[9]
)
def run_GFPGAN(*args, **kwargs):
time.sleep(.1)
return "yo"
def run_RealESRGAN(*args, **kwargs):
time.sleep(.2)
return "yo"
class model():
def __init__():
pass
class opt():
def __init__(self, name):
self.name = name
no_progressbar_hiding = True
css_hide_progressbar = """
.wrap .m-12 svg { display:none!important; }
.wrap .m-12::before { content:"Loading..." }
.progress-bar { display:none!important; }
.meta-text { display:none!important; }
"""
css = css_hide_progressbar
css = css + """
[data-testid="image"] {min-height: 512px !important};
#main_body {display:none !important};
#main_body>.col:nth-child(2){width:200%;}
"""
user_defaults = {}
# make sure these indicies line up at the top of txt2img()
txt2img_toggles = [
'Create prompt matrix (separate multiple prompts using |, and get all combinations of them)',
'Normalize Prompt Weights (ensure sum of weights add up to 1.0)',
'Save individual images',
'Save grid',
'Sort samples by prompt',
'Write sample info files',
'jpg samples',
]
if GFPGAN is not None:
txt2img_toggles.append('Fix faces using GFPGAN')
if RealESRGAN is not None:
txt2img_toggles.append('Upscale images using RealESRGAN')
txt2img_defaults = {
'prompt': '',
'ddim_steps': 50,
'toggles': [1, 2, 3],
'sampler_name': 'k_lms',
'ddim_eta': 0.0,
'n_iter': 1,
'batch_size': 1,
'cfg_scale': 7.5,
'seed': '',
'height': 512,
'width': 512,
'fp': None,
'submit_on_enter': 'Yes',
'variant_amount': 0,
'variant_seed': ''
}
if 'txt2img' in user_defaults:
txt2img_defaults.update(user_defaults['txt2img'])
txt2img_toggle_defaults = [txt2img_toggles[i] for i in txt2img_defaults['toggles']]
sample_img2img = "assets/stable-samples/img2img/sketch-mountains-input.jpg"
sample_img2img = sample_img2img if os.path.exists(sample_img2img) else None
imgproc_defaults = {
'prompt': '',
'ddim_steps': 50,
'sampler_name': 'k_lms',
'cfg_scale': 7.5,
'seed': '',
'height': 512,
'width': 512,
'denoising_strength': 0.30
}
imgproc_mode_toggles = [
'Fix Faces',
'Upscale'
]
# make sure these indicies line up at the top of img2img()
img2img_toggles = [
'Create prompt matrix (separate multiple prompts using |, and get all combinations of them)',
'Normalize Prompt Weights (ensure sum of weights add up to 1.0)',
'Loopback (use images from previous batch when creating next batch)',
'Random loopback seed',
'Save individual images',
'Save grid',
'Sort samples by prompt',
'Write sample info files',
'jpg samples',
]
if GFPGAN is not None:
img2img_toggles.append('Fix faces using GFPGAN')
if RealESRGAN is not None:
img2img_toggles.append('Upscale images using RealESRGAN')
img2img_mask_modes = [
"Keep masked area",
"Regenerate only masked area",
]
img2img_resize_modes = [
"Just resize",
"Crop and resize",
"Resize and fill",
]
img2img_defaults = {
'prompt': '',
'ddim_steps': 50,
'toggles': [1, 4, 5],
'sampler_name': 'k_lms',
'ddim_eta': 0.0,
'n_iter': 1,
'batch_size': 1,
'cfg_scale': 5.0,
'denoising_strength': 0.75,
'mask_mode': 0,
'resize_mode': 0,
'seed': '',
'height': 512,
'width': 512,
'fp': None,
}
if 'img2img' in user_defaults:
img2img_defaults.update(user_defaults['img2img'])
img2img_toggle_defaults = [img2img_toggles[i] for i in img2img_defaults['toggles']]
img2img_image_mode = 'sketch'
css_hide_progressbar = """
.wrap .m-12 svg { display:none!important; }
.wrap .m-12::before { content:"Loading..." }
.progress-bar { display:none!important; }
.meta-text { display:none!important; }
"""
styling = """
[data-testid="image"] {min-height: 512px !important}
* #body>.col:nth-child(2){width:250%;max-width:89vw}
#generate{width: 100%; }
#prompt_row input{
font-size:20px
}
input[type=number]:disabled { -moz-appearance: textfield;+ }
"""
demo = draw_gradio_ui(opt,
user_defaults=user_defaults,
txt2img=txt2img,
img2img=img2img,
txt2img_defaults=txt2img_defaults,
txt2img_toggles=txt2img_toggles,
txt2img_toggle_defaults=txt2img_toggle_defaults,
show_embeddings=hasattr(model, "embedding_manager"),
img2img_defaults=img2img_defaults,
img2img_toggles=img2img_toggles,
img2img_toggle_defaults=img2img_toggle_defaults,
img2img_mask_modes=img2img_mask_modes,
img2img_resize_modes=img2img_resize_modes,
sample_img2img=sample_img2img,
imgproc_defaults=imgproc_defaults,
imgproc_mode_toggles=imgproc_mode_toggles,
RealESRGAN=RealESRGAN,
GFPGAN=GFPGAN,
run_GFPGAN=run_GFPGAN,
run_RealESRGAN=run_RealESRGAN
)
# demo.queue()
demo.launch(share=False, debug=True)

1730
webui_streamlit.py
View File

File diff suppressed because it is too large Load Diff

100
webui_streamlit.yaml
View File

@ -1,100 +0,0 @@
# UI defaults configuration file. It is automatically loaded if located at configs/webui/webui_streamlit.yaml.
# Any changes made here will be available automatically on the web app without having to stop it.
general:
gpu: 0
outdir: outputs
ckpt: "models/ldm/stable-diffusion-v1/model.ckpt"
fp:
name: 'embeddings/alex/embeddings_gs-11000.pt'
GFPGAN_dir: "./src/gfpgan"
RealESRGAN_dir: "./src/realesrgan"
RealESRGAN_model: "RealESRGAN_x4plus"
outdir_txt2img: outputs/txt2img-samples
outdir_img2img: outputs/img2img-samples
gfpgan_cpu: False
esrgan_cpu: False
extra_models_cpu: False
extra_models_gpu: False
save_metadata: True
skip_grid: False
skip_save: False
grid_format: "jpg:95"
n_rows: -1
no_verify_input: False
no_half: False
precision: "autocast"
optimized: False
optimized_turbo: False
update_preview: True
update_preview_frequency: 1
txt2img:
prompt:
height: 512
width: 512
cfg_scale: 7.5
seed: ""
batch_count: 1
batch_size: 1
sampling_steps: 50
separate_prompts: False
normalize_prompt_weights: True
save_individual_images: True
save_grid: True
group_by_prompt: True
save_as_jpg: False
use_GFPGAN: True
use_RealESRGAN: True
RealESRGAN_model: "RealESRGAN_x4plus"
variant_amount: 0.0
img2img:
prompt:
sampling_steps: 50
# Adding an int to toggles enables the corresponding feature.
# 0: Create prompt matrix (separate multiple prompts using |, and get all combinations of them)
# 1: Normalize Prompt Weights (ensure sum of weights add up to 1.0)
# 2: Loopback (use images from previous batch when creating next batch)
# 3: Random loopback seed
# 4: Save individual images
# 5: Save grid
# 6: Sort samples by prompt
# 7: Write sample info files
# 8: jpg samples
# 9: Fix faces using GFPGAN
# 10: Upscale images using Real-ESRGAN
sampler_name: k_lms
denoising_strength: 0.75
# 0: Keep masked area
# 1: Regenerate only masked area
mask_mode: 0
# 0: Just resize
# 1: Crop and resize
# 2: Resize and fill
resize_mode: 0
# Leave blank for random seed:
seed:
ddim_eta: 0.0
cfg_scale: 5.0
batch_count: 1
batch_size: 1
height: 512
width: 512
# Textual inversion embeddings file path:
fp: ""
loopback: False
random_seed_loopback: False
separate_prompts: False
normalize_prompt_weights: True
save_individual_images: True
save_grid: True
group_by_prompt: True
save_as_jpg: False
use_GFPGAN: True
use_RealESRGAN: False
RealESRGAN_model: "RealESRGAN_x4plus"
variant_amount: 0.0
gfpgan:
strength: 100