IndicTransTokenizer

The goal of this repository is to provide a simple, modular, and extendable tokenizer for IndicTrans2 and be compatible with the HuggingFace models released.

Pre-requisites

Python 3.8+
Indic NLP Library
Other requirements as listed in requirements.txt

Configuration

Editable installation (Note, this may take a while):

git clone https://github.com/VarunGumma/IndicTransTokenizer
cd IndicTransTokenizer

pip install --editable ./

Usage

import torch
from transformers import AutoModelForSeq2SeqLM
from IndicTransTokenizer import IndicProcessor, IndicTransTokenizer

tokenizer = IndicTransTokenizer(direction="en-indic")
ip = IndicProcessor(inference=True)
model = AutoModelForSeq2SeqLM.from_pretrained("ai4bharat/indictrans2-en-indic-dist-200M", trust_remote_code=True)

sentences = [
    "This is a test sentence.",
    "This is another longer different test sentence.",
    "Please send an SMS to 9876543210 and an email on newemail123@xyz.com by 15th October, 2023.",
]

batch = ip.preprocess_batch(sentences, src_lang="eng_Latn", tgt_lang="hin_Deva")
batch = tokenizer(batch, src=True, return_tensors="pt")

with torch.inference_mode():
    outputs = model.generate(**batch, num_beams=5, num_return_sequences=1, max_length=256)

outputs = tokenizer.batch_decode(outputs, src=False)
outputs = ip.postprocess_batch(outputs, lang="hin_Deva")
print(outputs)

>>> ['यह एक परीक्षण वाक्य है।', 'यह एक और लंबा अलग परीक्षण वाक्य है।', 'कृपया 9876543210 पर एक एस. एम. एस. भेजें और 15 अक्टूबर, 2023 तक newemail123@xyz.com पर एक ईमेल भेजें।']

indic_evaluate function is a python implementation of compute_metrics.sh

from IndicTransTokenizer.evaluate import indic_evaluate

# this method returns a dictionary with BLEU and ChrF2++ scores with appropriate signatures
scores = indic_evaluate(tgt_lang=tgt_lang, preds=pred_file, refs=ref_file) 

# alternately, you can pass the list of predictions and references instead of files 
# scores = indic_evaluate(tgt_lang=tgt_lang, preds=preds, refs=refs)

For using the tokenizer to train/fine-tune the model, just set the inference argument of IndicProcessor to False.

Authors

Varun Gumma (varun230999@gmail.com)
Jay Gala (jaygala24@gmail.com)
Pranjal Agadh Chitale (pranjalchitale@gmail.com)
Raj Dabre (prajdabre@gmail.com)

Bugs and Contribution

Since this a bleeding-edge module, you may encounter broken stuff and import issues once in a while. In case you encounter any bugs or want additional functionalities, please feel free to raise Issues/Pull Requests or contact the authors.

Citation

If you use our codebase, models or tokenizer, please do cite the following paper:

@article{
    gala2023indictrans,
    title={IndicTrans2: Towards High-Quality and Accessible Machine Translation Models for all 22 Scheduled Indian Languages},
    author={Jay Gala and Pranjal A Chitale and A K Raghavan and Varun Gumma and Sumanth Doddapaneni and Aswanth Kumar M and Janki Atul Nawale and Anupama Sujatha and Ratish Puduppully and Vivek Raghavan and Pratyush Kumar and Mitesh M Khapra and Raj Dabre and Anoop Kunchukuttan},
    journal={Transactions on Machine Learning Research},
    issn={2835-8856},
    year={2023},
    url={https://openreview.net/forum?id=vfT4YuzAYA},
    note={}
}

Note

This tokenizer module is currently not compatible with the PreTrainedTokenizer module from HuggingFace. Hence, we are actively looking for Pull Requests to port this tokenizer to HF. Any leads on that front are welcome!

santoshdahal2016/IndicTransTokenizer