spartan-minhbui / sentiment-analysis

Pipeline for sentiment classification, use trainer of HuggingFace and use ONNX to export model

Finetune pretrained model for Sentiment analysis

Model Architecture

• Try some pretrained model:
  • PhoBERT: vinai/phobert-base-v2
  • Bloom: bigscience/bloom-560m
• Architecture: XXXXForSequenceClassification
• Use Trainer of Huggingface to training model

Dataset

• Dataset: 30K e-commerce reviews
• Labels:
  • NEG: Negative
  • POS: Positive
  • NEU: Neutral

Optimization

• Use some optimization techniques to optimize ONNX - Optim
• See: pipeline/onnx_converter.py

How to run

0. Note!!!

• Pass model_class to init class SentimentProcessor.
• Pass n_folds != None if you want to training with K-fold validation
• If you use Bloom, you should pass use_lora=True

1. Export environment variables: while read LINE; do export "$LINE"; done < .env
2. Run training: PRETRAINED_PATH=bigscience/bloom-560m CUDA_VISIBLE_DEVICES=1,2 python -m torch.distributed.launch --nproc_per_node 2 --master-port=30000 pipeline/trainer.py