rayxu14 / sl4du

Learning an Effective Context-Response Matching Model with Self-Supervised Tasks for Retrieval-based Dialogues

SL4DU

Environment

Option 1: container

frontlibrary/transformers-pytorch-gpu:4.6.1-pyarrow

~$ docker run --runtime=nvidia -it --rm -v $HOME/SL4DU:/workspace frontlibrary/transformers-pytorch-gpu:4.6.1-pyarrow

Option 2: build from scatch

• Python==3.9 (There may be some problem for numpy with 3.10)
• nltk
• numpy (can be automatically installed when installing scipy)
• scipy (if you have problem, see this solution)
• torch==1.8
• pyarrow
• tqdm
• transformers==4.5.1
• sklearn
• stop_words

Reproduce step: an example

1. Initialize directories

   SL4DU
       code
       data
       pretrained
   

2. Download code and the Ubuntu data

   ~/SL4DU/code$ git clone https://github.com/RayXu14/SL4DU.git
   ~/SL4DU/data$ wget https://www.dropbox.com/s/2fdn26rj6h9bpvl/ubuntu_data.zip
   ~/SL4DU/data$ unzip ubuntu_data.zip

3. Add bert-base-uncased pretrained model in pretrained
   • config.json
   • vocab.txt
   • pytorch_model.bin
4. Preprocess data

   ~/SL4DU/code/SL4DU$ python3 preprocess.py --task=RS --dataset=Ubuntu --raw_data_path=../../data/ubuntu_data --pkl_data_path=../../data/ubuntu_data --pretrained_model=bert-base-uncased

5. Reproduce BERT result

   ~/SL4DU/code/SL4DU$ python3 -u train.py --save_ckpt --task=RS --dataset=Ubuntu --pkl_data_path=../../data/ubuntu_data --pretrained_model=bert-base-uncased --add_EOT --freeze_layers=0 --train_batch_size=8 --eval_batch_size=100 --log_dir=? # --pkl_valid_file=test.pkl

6. Add post-ubuntu-bert-base-uncased in pretrained
   • Download whang's Ubuntu ckpt and use deprecated/whangpth2bin.py to transform it into our form; compared to bert-base-uncased, only need to +1 for vocab size in config.json and add a new word [EOS] after vocab.txt
   • Or use our pretrained models (already transformed) instead
7. Reproduce BERT-VFT result

   ~/SL4DU/code/SL4DU$ python3 -u train.py --save_ckpt --task=RS --dataset=Ubuntu --pkl_data_path=../../data/ubuntu_data --pretrained_model=post-ubuntu-bert-base-uncased --freeze_layers=8 --train_batch_size=16 --eval_batch_size=100 --log_dir=? #--pkl_valid_file=test.pkl

8. Reproduce SL4RS result

   ~/SL4DU/code/SL4DU$ python3 -u train.py --save_ckpt --task=RS --dataset=Ubuntu --pkl_data_path=../../data/ubuntu_data --pretrained_model=post-ubuntu-bert-base-uncased --freeze_layers=8 --train_batch_size=4 --eval_batch_size=100 --log_dir=? --use_NSP --use_UR --use_ID --use_CD --train_view_every=80 #--pkl_valid_file=test.pkl

9. Evaluation

   ~/SL4DU/code/SL4DU$ python3 -u eval.py --task=Ubuntu --data_path=../../data/ubuntu_data --pretrained_model=post-ubuntu-bert-base-uncased --freeze_layers=8 --eval_batch_size=100 --log_dir ? --load_path=?

Pretrained on yourself

Using Whang's repo

Remember to transform the saved model to our form using deprecated/whangpth2bin.py.

Additional information for pretraining settings

set the number of epochs as 2 for post-training with 10 duplication data and set the virtual batch size as 384