A tool to reduce the size of Hugging Face models via vocabulary trimming.
Reason this release was yanked:
Is buggy and does not install correctly.
Project description
hf-trim
A tool to reduce the size of Hugging Face models via vocabulary trimming.
The library currently supports the following models (and their pretrained versions available on the Hugging Face Models hub);
- BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation
- mBART: Multilingual Denoising Pre-training for Neural Machine Translation
- T5: Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
- mT5: A Massively Multilingual Pre-trained Text-to-Text Transformer
Installation
Run the following command to install from PyPI;
$ pip install hf-trim
You can also install from source;
$ git clone https://github.com/IamAdiSri/hf-trim
$ cd hf-trim
$ pip install .
Usage
Simple Example
from transformers import MT5Config, MT5Tokenizer, MT5ForConditionalGeneration
from hftrim.TokenizerTrimmer import TokenizerTrimmer
from hftrim.ModelTrimmer import ModelTrimmer
data = [
" UN Chief Says There Is No Military Solution in Syria",
"Şeful ONU declară că nu există o soluţie militară în Siria"
]
# load pretrained config, tokenizer and model
config = MT5Config.from_pretrained("google/mt5-small")
tokenizer = MT5Tokenizer.from_pretrained("google/mt5-small")
model = MT5ForConditionalGeneration.from_pretrained("google/mt5-small")
# trim tokenizer
tt = TokenizerTrimmer(tokenizer)
tt.make_vocab(data)
tt.make_tokenizer()
# trim model
mt = ModelTrimmer(model, config, tt.trimmed_tokenizer)
mt.make_weights(tt.trimmed_vocab_ids)
mt.make_model()
You can directly use the trimmed model with mt.trimmed_model
and the trimmed tokenizer with tt.trimmed_tokenizer
.
Saving and Loading
# save with
tt.trimmed_tokenizer.save_pretrained('trimT5')
mt.trimmed_model.save_pretrained('trimT5')
# load with
config = MT5Config.from_pretrained("trimT5")
tokenizer = MT5Tokenizer.from_pretrained("trimT5")
model = MT5ForConditionalGeneration.from_pretrained("trimT5")
Limitations
- Fast tokenizers are currently unsupported.
- Tensorflow and Flax models are currently unsupported.
Roadmap
- Add support for MarianMT models.
- Add support for FSMT models.
Issues
Feel free to open an issue if you run into bugs, have any queries or want to request support for an architecture.
Contributing
Contributions are welcome, especially those adding functionality for currently unsupported models.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.