Dialogue Summarization

Installation

pip install -r requirements.txt

Note: The current code base has been tested with 2 Nvidia 3090 GPUs. We also use wandb cloud-based logger which you may need to register and login first.

Data generation

Download aihub dialogue summary dataset.

python data_generator.py --path aihub_dialogue_summary --output data 

Update tokenizer for dialogue summary dataset

P01, P02와 같은 speaker id를 하나의 token으로 처리함.

1. Download gogamza/kobart-base-v1 tokenizer
2. Replace <unusedX> tokens into others

python update_tokenizer.py

Post-training

torchrun --nproc_per_node=2 run_post_train.py \
        --model_path "kobart-dialogue" \
        --model_name "gogamza/kobart-base-v1" \
        --run_name "kobart-post_train" \
        --do_train \
        --report_to "none" \
        --dialogue_max_seq_length 1024 \
        --train_file data/dialogue.json \
        --output_dir "checkpoints/kobart-post_train" \
        --learning_rate 0.0005 \
        --warmup_steps 50000 \
        --per_device_train_batch_size 8 \
        --max_steps 500000 \
        --save_strategy steps \
        --dataloader_num_workers 16 \
        --save_steps 10000 \
        --save_total_limit 3 \
        --gradient_accumulation_steps 1 \
        --logging_steps 100 

Fine-tuning

torchrun --nproc_per_node=1 run_summarization.py \
  --model_path "kobart-dialogue" \
  --model_name "gogamza/kobart-base-v1" \
  --run_name "kobart-dialsumm" \
  --do_train \
  --do_eval \
  --do_predict \
  --report_to "none" \
  --train_file data/train.csv \
  --valididation_file data/valid.csv \
  --predict_file data/predict.csv \
  --max_source_length 512 \
  --max_target_length 128 \
  --output_dir "checkpoints/kobart-dialsumm" \
  --learning_rate 5e-5 \
  --warmup_steps 50 \
  --per_device_train_batch_size 32 \
  --per_device_eval_batch_size 8 \
  --max_steps 1000 \
  --save_strategy steps \
  --evaluation_strategy steps \
  --dataloader_num_workers 10 \
  --save_steps 10 \
  --eval_steps 10 \
  --logging_steps 10 \
  --save_total_limit 3 \
  --load_best_model_at_end \
  --label_smoothing_factor 0.1 \
  --gradient_accumulation_steps 2 \
  --overwrite_cache \
  --fp16 \
  --predict_with_generate

Loss Graph

• Post train 과정
  • 400 ~ 600 step 범위에서 loss 수렴 (해당 부분에서 early stopping), 이 후 loss 상승 후 다시 수렴
  • Post-training 데이터 대략 28만개, 약 2 epoch 정도 최저점 부근에서 loss 수렴 하고, 이 후 loss 상승 (Overfitting 발생 추정)

Performance

Model	R1	R2	RL
Naive Fine tuning	31.3613	17.6152	28.2803
Post training then Fine tuning	32.5910	18.5439	29.4671