/CALM-Dialogue

Official code for the paper "Context-Aware Language Modeling for Goal-Oriented Dialogue Systems"

Primary LanguagePythonMIT LicenseMIT

Context-Aware Language Modeling for Goal-Oriented Dialogue Systems

Official code for the paper "Context-Aware Language Modeling for Goal-Oriented Dialogue Systems"

project site | arxiv

setup

  1. install requirements
    • conda create -n calm --file requirements.txt
    • pip install -r pip_requirements.txt
    • conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
  2. export PYTHONPATH="$PWD/offline_airdialogue"
  3. Download the processed data and model checkpoints here. The outputs/ folder contains checkpoints for our main model, our task pretrained model, and our customer bot.

Training

(Note: all training runs use wandb by default, you can turn off wandb syncing in the config.)

  • cd scripts/train

  • To run data-parallel multi-GPU training, on any of the commands below replace python <script_path> with python -m torch.distributed.launch --nproc_per_node <n_GPUs> --use_env <script_path>.

  • Pretraining CALM
    (two variants of the auxiliary loss function)

    • script: python train_pretrain_table_agent.py
      config: config/train_pretrain_table_agent.yaml

    • script: python train_pretrain_simplified_aux_gpt2.py
      config: config/train_pretrain_simplified_aux_gpt2.yaml

  • Training the customer bot

    • script: python train_customer.py
      config: config/train_customer_bot.yaml

  • Training CALM
    (two variants of the auxiliary loss function)

    • script: python train_real_table_agent.py
      config: config/train_real_table_agent.yaml

    • script: python train_simplified_aux_gpt2.py
      config: config/train_simplified_aux_agent.yaml

  • Training Standard LM

    • script: python train_basic_agent.py
      config: config/train_basic_agent.yaml

  • Training the reward model for Model Based Rollout Planning

    • script: python train_constraint_parser.py
      config: config/train_constraint_parser.yaml

Evaluating

  • cd scripts/eval

  • Simulated Evaluation

    • script: python selfplay_eval.py
      config: config/selfplay_eval.yaml

      • A log of results will be saved to the location specified by selfplay/outputs_file in the config. To print out the success rate for the selfplay run: python compute_results.py --results_file <your_eval_outputs_file>

      • Note: selfplay evaluation will by default use all the GPUs available on your machine. To Specify which GPUs to use, prefix the command with CUDA_VISIBLE_DEVICES=<comma_seperated_list_of_gpu_indicies>

  • Language Quality Evaluation

    • script: python language_quality_eval.py
      config: config/language_eval.yaml

      • To parallelize evaluation across on multiple-GPUs, run: python -m torch.distributed.launch --nproc_per_node <n_GPUs> --use_env language_quality_eval.py