guillaume-be/rust-bert
1
1
Fork 0
mirror of https://github.com/guillaume-be/rust-bert.git synced 2024-08-17 08:30:25 +03:00
Code Issues Packages Projects Releases Wiki Activity
1,109 Commits 5 Branches 0 Tags 5.8 MiB
main
Commit Graph

14 Commits

Author SHA1 Message Date
Abdulrhman Alkhodiry
33b2944298
Update dependencies and fix convert_model (#458) 
* feat: Update dependencies in Cargo.toml

Update the dependencies in Cargo.toml to their latest versions:
- rust_tokenizers: 8.1.1
- tch: 0.16.0 (with features = ["download-libtorch"])
- serde_json: 1
- serde: 1 (with features = ["derive"])
- ordered-float: 4.2.0
- uuid: 1 (with features = ["v4"])
- thiserror: 1
- half: 2
- regex: 1.6
- cached-path: 0.6 (with default-features = false and optional = true)
- dirs: 5 (optional = true)
- lazy_static: 1 (optional = true)
- ort: 1.16.3 (optional = true, default-features = false, features = ["half"])
- ndarray: 0.15 (optional = true)
- tokenizers: 0.19.1 (optional = true, default-features = false, features = ["onig"])

```

* chore: Update .gitignore and requirements.txt, and improve convert_model.py

Update .gitignore to exclude the /models/ and /.venv/ directories, and the convert_model.log file.

Remove the requirements.txt file.

In convert_model.py:
- Add a new function, `zipfile_factory`, to handle zip file creation.
- Update the logger configuration to log debug messages to a file named `convert_model.log`.

* delete duplicate requirements file

* update CI req file path

* missing requests dependency

---------

Co-authored-by: Abdulrhman Alkhodiry <aalkhodiry@jahez.net>
Co-authored-by: Guillaume Becquin <guillaume.becquin@gmail.com>
 2024-06-30 08:41:10 +01:00
guillaume-be
107fb214d9
Support wildcard for source files in convert utilities (#399) 
* Support wildcard for source files

* - Support offloading at end of every file for conversion
 2023-07-02 08:36:21 +01:00
guillaume-be
a74d023583
Add bfloat16 tensor loading support (#396) 2023-06-25 09:21:52 +01:00
Joseph Hajduk
2bff63b2ee
updated tch-rs to 0.13.0 (#380) 
* updated tch-rs to 0.13.0
find replaced of_slice to from_slice as per
008fff6cc0/CHANGELOG.md

* fixed formatting

* Add download feature and update CI

* add build script, update CI

* updated chanelog, readme, convert script

* fixed wrong position for build script

* added libtorch download to dependencies download test script

* args reordering

---------

Co-authored-by: josephhajduk <joseph@solidys.dev>
Co-authored-by: Guillaume Becquin <guillaume.becquin@gmail.com>
 2023-05-21 18:41:18 +01:00
Panakotta00
431c3ef11b
convert_model.py add support for multiple binary files (#347) 2023-04-06 21:32:51 +01:00
Romain Leroux
c448862185
Add GPT-J support (#285) (#288) 
* Add GPT-J support (#285)

* Improve GPT-J implementation

* Improve GPT-J tests

* Adapt GPT-J to latest master branch

* Specify how to convert GPT-J weights instead of providing them
 2023-02-15 19:10:47 +00:00
guillaume-be
d7e9c03694
Long t5 implementation (#333) 
* LongT5 config implementation

* LongT5 WiP: utility functions 1

* LongT5 WiP: utility functions (2)

* LongT5 WiP: utility functions (3)

* LongT5 WiP: utility functions (4)

* made T5 FF activations generic, expose T5 modules to crate

* Longt% local attention WIP

* LongT5 local attention

* LongT5 global attention WIP

* LongT5 global attention

* LongT5 attention modules (WIP)

* align LongT5 position bias with T5

* Addition of LongT5Block

* LongT5Stack WiP

* LongT5Stack implementation

* LongT5Model implementation

* LongT5ForConditionalGeneration implementation

* Addition of LongT5Generator, inclusion in pipelines

* LongT5 attention fixes

* Fix MIN/MAX dtype computation, mask for longt5

* Updated min/max and infinity computation across models

* GlobalTransient attention fixes

* Updated changelog, readme, tests, clippy
 2023-02-12 16:18:20 +00:00
Romain Leroux
4d8a298586
Add sbert implementation for inference (#250) 
* Add sbert implementation for inference

* Fix clippy warnings

* Refactor sentence embeddings into a dedicated pipeline

* Add output_attentions and output_hidden_states to T5Config

* Add sbert implementation for inference

* Fix clippy warnings

* Refactor sentence embeddings into a dedicated pipeline

* Add output_attentions and output_hidden_states to T5Config

* Improve sentence_embeddings implementation

* Dedicated tokenizer config for strip_accents and add_prefix_space

* Rename forward to encode_as_tensor

* Remove _conf from Dense layer

* Add sentence embeddings docs

* Addition of remote resources and tests update

* Merge feature branch and fix doctests

* Add SentenceEmbeddingsBuilder<Remote> and improve remote resources

* Use tch::no_grad in sentence embeddings

* Updated changelog, registration of sentence embeddings integration tests

Co-authored-by: Guillaume Becquin <guillaume.becquin@gmail.com>
 2022-06-21 20:24:09 +01:00
Guillaume Becquin
60591c0644 Updated DeBERTa configuration parsing 2021-11-28 12:43:35 +01:00
guillaume-be
4282d7b5c4
Reuse gpt2 embeddings (#160) 
* Updated GPT2 to re-use embeddings for LM head

* Updated conversion utilities

* Updated changelog
 2021-06-12 11:11:34 +02:00
Guillaume B
b6db7cacfb MBart validation, weights updated 2021-06-05 11:47:56 +02:00
Guillaume B
05295068b8 Updated README (addition of benchmarks and file conversion) 2021-02-02 17:51:39 +01:00
Guillaume B
f8ed5079c1 initial commit for ProphetNet 2021-01-07 10:41:50 +01:00
Guillaume B
dca5252f38 Addition of generic conversion script 2020-12-04 16:35:13 +01:00