Paraphrasing with T5

• This model generates and output sentence that preserves the meaning of input sentence with variations in word choise and grammar.

• This repo uses the dataset humarin/chatgpt-paraphrases from huggingface. You can use other datasets with some modification.

• This repo uses model t5-small - small version of T5 model.

Setup

• To create Docker environment:

docker build -t paraphrasing-env .
docker run -itd --gpus all --restart always -v $(pwd)/:/workspace --name t5-env paraphrasing-env:latest
docker exec -it text-env bash

Train

• To train:

python3 main.py \ 
--scenario train \
--num_epochs 30 \
--batch_size 16 \
--save_total_limit 3

• Flag:
  • --save_total_limit: number of times to save checkpoints.

Inference

• To infer:

python3 main.py \
--scenario infer \
--sentence "What are the best place in Vietnam?" \
--num_return_sequences 5 \
--max_length 256

• Flag:
  • --max_length: maxium length of sentence.

Triton Deployment

• To deploy with Triton Server: https://github.com/trinhtuanvubk/t5-triton-pipeline