Optimizing Adapters for Neural Machine Translation

This repositary hosts my experiments for the project, I did with OffNote Labs. Blog post summarizing my project can be found here.

Details

mBART code is taken from 🤗Transformers & I modified it to add adapters. I made changes to only this folder from 🤗Transformers. Rest of 🤗Transformers code is untouched.

Adapters specific code can be found here. Other boilerplate code can be found here.

Checkpoints

Model details	Checkpoint	Size
hin->eng adapters	adapters-hin-eng	104.6 MB
hin->eng Fine-tuned model	mbart-bhasha-hin-eng	2.3 GB
guj->eng adapters	adapters-guj-eng	104.6 MB
guj->eng Fine-tuned model	mbart-bhasha-guj-eng	2.3 GB

Note: Bhasha guj-eng dataset has only 59K samples; while hin-eng dataset has 260K samples

Other checkpoints can be found here.

Setting Up

# install requirements
pip3 install -r requirements.txt
cd src

# downloading & unzipping dataset
wget http://preon.iiit.ac.in/~jerin/resources/datasets/pib_v1.3.tar.gz
unzip pib_v1.3.tar.gz

# Adapters checkpoints (it's just 200 MB) can be downloaded using:
git clone https://huggingface.co/vasudevgupta/offnote-mbart-adapters-bhasha

Using this project

# run this from src/ directory
from transformers import MBartForConditionalGeneration, MBartTokenizer
from transformers.models.bart.adapter_modeling import AdapterConfig

# initialize mBART from pre-trained weights
tokenizer = MBartTokenizer.from_pretrained("facebook/mbart-large-cc25")
model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-cc25")

# Now let's add adapters which will enable multilingual translation

# deciding what all adapters to add
adapter_args = {
    "dec_ffn_adapter": True,
    "enc_self_attn_adapter": True,
    "enc_tok_embed_adapter": True,
    "dec_tok_embed_adapter": True,
    "enc_self_attn_adapter_config": AdapterConfig(input_size=1024),
    "dec_ffn_adapter_config": AdapterConfig(input_size=1024),
    "enc_tok_embed_adapter_config": AdapterConfig(input_size=1024),
    "dec_tok_embed_adapter_config": AdapterConfig(input_size=1024),
}

# add randomly-initialized adapters
model.add_adapter_(**adapter_args)

# loading hin-eng adapters
model.load_adapter("offnote-mbart-adapters-bhasha/adapters-hin-eng.pt", map_location="cpu")

# loading guj-eng adapters
model.load_adapter("offnote-mbart-adapters-bhasha/adapter-guj-eng.pt", map_location="cpu")

# model is ready for inference just like any other 🤗Transformers

Reproducing our results

# for normal fine-tuning
python3 train.py --config "main"

# for training with best-adapters case on hin-eng
python3 train.py --config "best_adapters_hin"
# for normal-finetuning on hin-eng
python3 train.py --config "full_train_hin"

# for training with best-adapters case on guj-eng
python3 train.py --config "best_adapters_guj"
# for normal-finetuning on guj-eng
python3 train.py --config "full_train_guj"

# for inference & bleu score logging over test, run this (after changing config)
python3 infer.py