stable-diffusion-webui/modules/textual_inversion/textual_inversion.py

import os
import sys
import traceback

import torch
import tqdm
import html
import datetime
import csv

from PIL import Image, PngImagePlugin

from modules import shared, devices, sd_hijack, processing, sd_models, images, sd_samplers
import modules.textual_inversion.dataset
from modules.textual_inversion.learn_schedule import LearnRateScheduler

from modules.textual_inversion.image_embedding import (embedding_to_b64, embedding_from_b64,
                                                       insert_image_data_embed, extract_image_data_embed,
                                                       caption_image_overlay)

class Embedding:
    def __init__(self, vec, name, step=None):
        self.vec = vec
        self.name = name
        self.step = step
        self.shape = None
        self.vectors = 0
        self.cached_checksum = None
        self.sd_checkpoint = None
        self.sd_checkpoint_name = None
        self.optimizer_state_dict = None

    def save(self, filename):
        embedding_data = {
            "string_to_token": {"*": 265},
            "string_to_param": {"*": self.vec},
            "name": self.name,
            "step": self.step,
            "sd_checkpoint": self.sd_checkpoint,
            "sd_checkpoint_name": self.sd_checkpoint_name,
        }

        torch.save(embedding_data, filename)

        if shared.opts.save_optimizer_state and self.optimizer_state_dict is not None:
            optimizer_saved_dict = {
                'hash': self.checksum(),
                'optimizer_state_dict': self.optimizer_state_dict,
            }
            torch.save(optimizer_saved_dict, filename + '.optim')

    def checksum(self):
        if self.cached_checksum is not None:
            return self.cached_checksum

        def const_hash(a):
            r = 0
            for v in a:
                r = (r * 281 ^ int(v) * 997) & 0xFFFFFFFF
            return r

        self.cached_checksum = f'{const_hash(self.vec.reshape(-1) * 100) & 0xffff:04x}'
        return self.cached_checksum


class EmbeddingDatabase:
    def __init__(self, embeddings_dir):
        self.ids_lookup = {}
        self.word_embeddings = {}
        self.skipped_embeddings = {}
        self.dir_mtime = None
        self.embeddings_dir = embeddings_dir
        self.expected_shape = -1

    def register_embedding(self, embedding, model):

        self.word_embeddings[embedding.name] = embedding

        # TODO changing between clip and open clip changes tokenization, which will cause embeddings to stop working
        ids = model.cond_stage_model.tokenize([embedding.name])[0]

        first_id = ids[0]
        if first_id not in self.ids_lookup:
            self.ids_lookup[first_id] = []

        self.ids_lookup[first_id] = sorted(self.ids_lookup[first_id] + [(ids, embedding)], key=lambda x: len(x[0]), reverse=True)

        return embedding

    def get_expected_shape(self):
        vec = shared.sd_model.cond_stage_model.encode_embedding_init_text(",", 1)
        return vec.shape[1]

    def load_textual_inversion_embeddings(self, force_reload = False):
        mt = os.path.getmtime(self.embeddings_dir)
        if not force_reload and self.dir_mtime is not None and mt <= self.dir_mtime:
            return

        self.dir_mtime = mt
        self.ids_lookup.clear()
        self.word_embeddings.clear()
        self.skipped_embeddings.clear()
        self.expected_shape = self.get_expected_shape()

        def process_file(path, filename):
            name, ext = os.path.splitext(filename)
            ext = ext.upper()

            if ext in ['.PNG', '.WEBP', '.JXL', '.AVIF']:
                embed_image = Image.open(path)
                if hasattr(embed_image, 'text') and 'sd-ti-embedding' in embed_image.text:
                    data = embedding_from_b64(embed_image.text['sd-ti-embedding'])
                    name = data.get('name', name)
                else:
                    data = extract_image_data_embed(embed_image)
                    name = data.get('name', name)
            elif ext in ['.BIN', '.PT']:
                data = torch.load(path, map_location="cpu")
            else:
                return

            # textual inversion embeddings
            if 'string_to_param' in data:
                param_dict = data['string_to_param']
                if hasattr(param_dict, '_parameters'):
                    param_dict = getattr(param_dict, '_parameters')  # fix for torch 1.12.1 loading saved file from torch 1.11
                assert len(param_dict) == 1, 'embedding file has multiple terms in it'
                emb = next(iter(param_dict.items()))[1]
            # diffuser concepts
            elif type(data) == dict and type(next(iter(data.values()))) == torch.Tensor:
                assert len(data.keys()) == 1, 'embedding file has multiple terms in it'

                emb = next(iter(data.values()))
                if len(emb.shape) == 1:
                    emb = emb.unsqueeze(0)
            else:
                raise Exception(f"Couldn't identify {filename} as neither textual inversion embedding nor diffuser concept.")

            vec = emb.detach().to(devices.device, dtype=torch.float32)
            embedding = Embedding(vec, name)
            embedding.step = data.get('step', None)
            embedding.sd_checkpoint = data.get('sd_checkpoint', None)
            embedding.sd_checkpoint_name = data.get('sd_checkpoint_name', None)
            embedding.vectors = vec.shape[0]
            embedding.shape = vec.shape[-1]

            if self.expected_shape == -1 or self.expected_shape == embedding.shape:
                self.register_embedding(embedding, shared.sd_model)
            else:
                self.skipped_embeddings[name] = embedding

        for fn in os.listdir(self.embeddings_dir):
            try:
                fullfn = os.path.join(self.embeddings_dir, fn)

                if os.stat(fullfn).st_size == 0:
                    continue

                process_file(fullfn, fn)
            except Exception:
                print(f"Error loading embedding {fn}:", file=sys.stderr)
                print(traceback.format_exc(), file=sys.stderr)
                continue

        print(f"Textual inversion embeddings loaded({len(self.word_embeddings)}): {', '.join(self.word_embeddings.keys())}")
        if len(self.skipped_embeddings) > 0:
            print(f"Textual inversion embeddings skipped({len(self.skipped_embeddings)}): {', '.join(self.skipped_embeddings.keys())}")

    def find_embedding_at_position(self, tokens, offset):
        token = tokens[offset]
        possible_matches = self.ids_lookup.get(token, None)

        if possible_matches is None:
            return None, None

        for ids, embedding in possible_matches:
            if tokens[offset:offset + len(ids)] == ids:
                return embedding, len(ids)

        return None, None


def create_embedding(name, num_vectors_per_token, overwrite_old, init_text='*'):
    cond_model = shared.sd_model.cond_stage_model

    with devices.autocast():
        cond_model([""])  # will send cond model to GPU if lowvram/medvram is active

    embedded = cond_model.encode_embedding_init_text(init_text, num_vectors_per_token)
    vec = torch.zeros((num_vectors_per_token, embedded.shape[1]), device=devices.device)

    for i in range(num_vectors_per_token):
        vec[i] = embedded[i * int(embedded.shape[0]) // num_vectors_per_token]

    # Remove illegal characters from name.
    name = "".join( x for x in name if (x.isalnum() or x in "._- "))
    fn = os.path.join(shared.cmd_opts.embeddings_dir, f"{name}.pt")
    if not overwrite_old:
        assert not os.path.exists(fn), f"file {fn} already exists"

    embedding = Embedding(vec, name)
    embedding.step = 0
    embedding.save(fn)

    return fn


def write_loss(log_directory, filename, step, epoch_len, values):
    if shared.opts.training_write_csv_every == 0:
        return

    if step % shared.opts.training_write_csv_every != 0:
        return
    write_csv_header = False if os.path.exists(os.path.join(log_directory, filename)) else True

    with open(os.path.join(log_directory, filename), "a+", newline='') as fout:
        csv_writer = csv.DictWriter(fout, fieldnames=["step", "epoch", "epoch_step", *(values.keys())])

        if write_csv_header:
            csv_writer.writeheader()

        epoch = (step - 1) // epoch_len
        epoch_step = (step - 1) % epoch_len

        csv_writer.writerow({
            "step": step,
            "epoch": epoch,
            "epoch_step": epoch_step,
            **values,
        })

def validate_train_inputs(model_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_model_every, create_image_every, log_directory, name="embedding"):
    assert model_name, f"{name} not selected"
    assert learn_rate, "Learning rate is empty or 0"
    assert isinstance(batch_size, int), "Batch size must be integer"
    assert batch_size > 0, "Batch size must be positive"
    assert isinstance(gradient_step, int), "Gradient accumulation step must be integer"
    assert gradient_step > 0, "Gradient accumulation step must be positive"
    assert data_root, "Dataset directory is empty"
    assert os.path.isdir(data_root), "Dataset directory doesn't exist"
    assert os.listdir(data_root), "Dataset directory is empty"
    assert template_file, "Prompt template file is empty"
    assert os.path.isfile(template_file), "Prompt template file doesn't exist"
    assert steps, "Max steps is empty or 0"
    assert isinstance(steps, int), "Max steps must be integer"
    assert steps > 0 , "Max steps must be positive"
    assert isinstance(save_model_every, int), "Save {name} must be integer"
    assert save_model_every >= 0 , "Save {name} must be positive or 0"
    assert isinstance(create_image_every, int), "Create image must be integer"
    assert create_image_every >= 0 , "Create image must be positive or 0"
    if save_model_every or create_image_every:
        assert log_directory, "Log directory is empty"

def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, clip_grad_mode, clip_grad_value, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
    save_embedding_every = save_embedding_every or 0
    create_image_every = create_image_every or 0
    validate_train_inputs(embedding_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding")

    shared.state.job = "train-embedding"
    shared.state.textinfo = "Initializing textual inversion training..."
    shared.state.job_count = steps

    filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt')

    log_directory = os.path.join(log_directory, datetime.datetime.now().strftime("%Y-%m-%d"), embedding_name)
    unload = shared.opts.unload_models_when_training

    if save_embedding_every > 0:
        embedding_dir = os.path.join(log_directory, "embeddings")
        os.makedirs(embedding_dir, exist_ok=True)
    else:
        embedding_dir = None

    if create_image_every > 0:
        images_dir = os.path.join(log_directory, "images")
        os.makedirs(images_dir, exist_ok=True)
    else:
        images_dir = None

    if create_image_every > 0 and save_image_with_stored_embedding:
        images_embeds_dir = os.path.join(log_directory, "image_embeddings")
        os.makedirs(images_embeds_dir, exist_ok=True)
    else:
        images_embeds_dir = None

    hijack = sd_hijack.model_hijack

    embedding = hijack.embedding_db.word_embeddings[embedding_name]
    checkpoint = sd_models.select_checkpoint()

    initial_step = embedding.step or 0
    if initial_step >= steps:
        shared.state.textinfo = "Model has already been trained beyond specified max steps"
        return embedding, filename
    scheduler = LearnRateScheduler(learn_rate, steps, initial_step)

    clip_grad = torch.nn.utils.clip_grad_value_ if clip_grad_mode == "value" else \
        torch.nn.utils.clip_grad_norm_ if clip_grad_mode == "norm" else \
        None
    if clip_grad:
        clip_grad_sched = LearnRateScheduler(clip_grad_value, steps, ititial_step, verbose=False)
    # dataset loading may take a while, so input validations and early returns should be done before this
    shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."
    old_parallel_processing_allowed = shared.parallel_processing_allowed

    pin_memory = shared.opts.pin_memory

    ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method)

    latent_sampling_method = ds.latent_sampling_method

    dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, latent_sampling_method=latent_sampling_method, batch_size=ds.batch_size, pin_memory=pin_memory)

    if unload:
        shared.parallel_processing_allowed = False
        shared.sd_model.first_stage_model.to(devices.cpu)

    embedding.vec.requires_grad = True
    optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate, weight_decay=0.0)
    if shared.opts.save_optimizer_state:
        optimizer_state_dict = None
        if os.path.exists(filename + '.optim'):
            optimizer_saved_dict = torch.load(filename + '.optim', map_location='cpu')
            if embedding.checksum() == optimizer_saved_dict.get('hash', None):
                optimizer_state_dict = optimizer_saved_dict.get('optimizer_state_dict', None)
    
        if optimizer_state_dict is not None:
            optimizer.load_state_dict(optimizer_state_dict)
            print("Loaded existing optimizer from checkpoint")
        else:
            print("No saved optimizer exists in checkpoint")

    scaler = torch.cuda.amp.GradScaler()

    batch_size = ds.batch_size
    gradient_step = ds.gradient_step
    # n steps = batch_size * gradient_step * n image processed
    steps_per_epoch = len(ds) // batch_size // gradient_step
    max_steps_per_epoch = len(ds) // batch_size - (len(ds) // batch_size) % gradient_step
    loss_step = 0
    _loss_step = 0 #internal

    last_saved_file = "<none>"
    last_saved_image = "<none>"
    forced_filename = "<none>"
    embedding_yet_to_be_embedded = False

    is_training_inpainting_model = shared.sd_model.model.conditioning_key in {'hybrid', 'concat'}
    img_c = None

    pbar = tqdm.tqdm(total=steps - initial_step)
    try:
        for i in range((steps-initial_step) * gradient_step):
            if scheduler.finished:
                break
            if shared.state.interrupted:
                break
            for j, batch in enumerate(dl):
                # works as a drop_last=True for gradient accumulation
                if j == max_steps_per_epoch:
                    break
                scheduler.apply(optimizer, embedding.step)
                if scheduler.finished:
                    break
                if shared.state.interrupted:
                    break

                if clip_grad:
                    clip_grad_sched.step(embedding.step)
            
                with devices.autocast():
                    x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
                    c = shared.sd_model.cond_stage_model(batch.cond_text)

                    if is_training_inpainting_model:
                        if img_c is None:
                            img_c = processing.txt2img_image_conditioning(shared.sd_model, c, training_width, training_height)

                        cond = {"c_concat": [img_c], "c_crossattn": [c]}
                    else:
                        cond = c

                    loss = shared.sd_model(x, cond)[0] / gradient_step
                    del x

                    _loss_step += loss.item()
                scaler.scale(loss).backward()

                # go back until we reach gradient accumulation steps
                if (j + 1) % gradient_step != 0:
                    continue
                
                if clip_grad:
                    clip_grad(embedding.vec, clip_grad_sched.learn_rate)

                scaler.step(optimizer)
                scaler.update()
                embedding.step += 1
                pbar.update()
                optimizer.zero_grad(set_to_none=True)
                loss_step = _loss_step
                _loss_step = 0

                steps_done = embedding.step + 1

                epoch_num = embedding.step // steps_per_epoch
                epoch_step = embedding.step % steps_per_epoch

                pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}]loss: {loss_step:.7f}")
                if embedding_dir is not None and steps_done % save_embedding_every == 0:
                    # Before saving, change name to match current checkpoint.
                    embedding_name_every = f'{embedding_name}-{steps_done}'
                    last_saved_file = os.path.join(embedding_dir, f'{embedding_name_every}.pt')
                    save_embedding(embedding, optimizer, checkpoint, embedding_name_every, last_saved_file, remove_cached_checksum=True)
                    embedding_yet_to_be_embedded = True

                write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, steps_per_epoch, {
                    "loss": f"{loss_step:.7f}",
                    "learn_rate": scheduler.learn_rate
                })

                if images_dir is not None and steps_done % create_image_every == 0:
                    forced_filename = f'{embedding_name}-{steps_done}'
                    last_saved_image = os.path.join(images_dir, forced_filename)

                    shared.sd_model.first_stage_model.to(devices.device)

                    p = processing.StableDiffusionProcessingTxt2Img(
                        sd_model=shared.sd_model,
                        do_not_save_grid=True,
                        do_not_save_samples=True,
                        do_not_reload_embeddings=True,
                    )

                    if preview_from_txt2img:
                        p.prompt = preview_prompt
                        p.negative_prompt = preview_negative_prompt
                        p.steps = preview_steps
                        p.sampler_name = sd_samplers.samplers[preview_sampler_index].name
                        p.cfg_scale = preview_cfg_scale
                        p.seed = preview_seed
                        p.width = preview_width
                        p.height = preview_height
                    else:
                        p.prompt = batch.cond_text[0]
                        p.steps = 20
                        p.width = training_width
                        p.height = training_height

                    preview_text = p.prompt

                    processed = processing.process_images(p)
                    image = processed.images[0] if len(processed.images) > 0 else None

                    if unload:
                        shared.sd_model.first_stage_model.to(devices.cpu)

                    if image is not None:
                        shared.state.current_image = image
                        last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
                        last_saved_image += f", prompt: {preview_text}"

                    if save_image_with_stored_embedding and os.path.exists(last_saved_file) and embedding_yet_to_be_embedded:

                        last_saved_image_chunks = os.path.join(images_embeds_dir, f'{embedding_name}-{steps_done}.png')

                        info = PngImagePlugin.PngInfo()
                        data = torch.load(last_saved_file)
                        info.add_text("sd-ti-embedding", embedding_to_b64(data))

                        title = "<{}>".format(data.get('name', '???'))

                        try:
                            vectorSize = list(data['string_to_param'].values())[0].shape[0]
                        except Exception as e:
                            vectorSize = '?'

                        checkpoint = sd_models.select_checkpoint()
                        footer_left = checkpoint.model_name
                        footer_mid = '[{}]'.format(checkpoint.hash)
                        footer_right = '{}v {}s'.format(vectorSize, steps_done)

                        captioned_image = caption_image_overlay(image, title, footer_left, footer_mid, footer_right)
                        captioned_image = insert_image_data_embed(captioned_image, data)

                        captioned_image.save(last_saved_image_chunks, "PNG", pnginfo=info)
                        embedding_yet_to_be_embedded = False

                    last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
                    last_saved_image += f", prompt: {preview_text}"

                shared.state.job_no = embedding.step

                shared.state.textinfo = f"""
<p>
Loss: {loss_step:.7f}<br/>
Step: {steps_done}<br/>
Last prompt: {html.escape(batch.cond_text[0])}<br/>
Last saved embedding: {html.escape(last_saved_file)}<br/>
Last saved image: {html.escape(last_saved_image)}<br/>
</p>
"""
        filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt')
        save_embedding(embedding, optimizer, checkpoint, embedding_name, filename, remove_cached_checksum=True)
    except Exception:
        print(traceback.format_exc(), file=sys.stderr)
        pass
    finally:
        pbar.leave = False
        pbar.close()
        shared.sd_model.first_stage_model.to(devices.device)
        shared.parallel_processing_allowed = old_parallel_processing_allowed

    return embedding, filename

def save_embedding(embedding, optimizer, checkpoint, embedding_name, filename, remove_cached_checksum=True):
    old_embedding_name = embedding.name
    old_sd_checkpoint = embedding.sd_checkpoint if hasattr(embedding, "sd_checkpoint") else None
    old_sd_checkpoint_name = embedding.sd_checkpoint_name if hasattr(embedding, "sd_checkpoint_name") else None
    old_cached_checksum = embedding.cached_checksum if hasattr(embedding, "cached_checksum") else None
    try:
        embedding.sd_checkpoint = checkpoint.hash
        embedding.sd_checkpoint_name = checkpoint.model_name
        if remove_cached_checksum:
            embedding.cached_checksum = None
        embedding.name = embedding_name
        embedding.optimizer_state_dict = optimizer.state_dict()
        embedding.save(filename)
    except:
        embedding.sd_checkpoint = old_sd_checkpoint
        embedding.sd_checkpoint_name = old_sd_checkpoint_name
        embedding.name = old_embedding_name
        embedding.cached_checksum = old_cached_checksum
        raise