Response Generation

Scripts to train Seq2Seq and Transformer models on the Amazon Topical-Chat Corpus. This code serves as the baseline for DSTC9 Track 3.

To train: python3 train.py --use_knowledge --transformer --save_path transformer/

To test: python3 test.py --use_knowledge --transformer --save_path transformer/

To serve interactive model with TF-IDF based fact selection: python3 dynamic.py --use_knowledge --transformer --save_path transformer/

Data

The pre-processed data can be found in data.zip. If you would like to use a different pre-processing strategy, please download the original data from here.

Pre-trained models

The pre-trained models can be found at: https://drive.google.com/file/d/1fPB45RDs_BcJ8KZeYQiauK3W1RsdY2hM/view?usp=sharing

python3 test.py --save_path /docker-share/data/transformer_trained --epoch 20
test_freq Epoch: 20 PPL: 18.540181650272256
test_rare Epoch: 20 PPL: 26.536356859318733

Contact

If you experience any issues with this code, please contact me at [email protected]

Results

• train.py

python3 train.py --use_knowledge --transformer --batch_size=8 --save_path transformer/
...
Epoch 20/20 Batch 22450/22464 Avg Loss 3.81 LR 0.0002                                                                                                                                                                                         
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 22464/22464 [35:46<00:00, 10.47it/s]

• test.py

CUDA_VISIBLE_DEVICES=1 python3 test.py --save_path transformer-2 --epoch 0                                                                                  
test_freq Epoch: 0 PPL: 13.543523391295622                  
test_rare Epoch: 0 PPL: 16.720335006129222 
CUDA_VISIBLE_DEVICES=1 python3 test.py --save_path transformer-2 --epoch 2
test_freq Epoch: 2 PPL: 6.41167198939869
test_rare Epoch: 2 PPL: 7.645547600879482