https://github.com/sea-snell/calm-dialogue

Official code for the paper "Context-Aware Language Modeling for Goal-Oriented Dialogue Systems"
https://github.com/sea-snell/calm-dialogue

deep-learning language-model nlp python pytorch reinforcement-learning

Last synced: 9 months ago
JSON representation

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

• Host: GitHub
• URL: https://github.com/sea-snell/calm-dialogue
• Owner: Sea-Snell
• License: mit
• Created: 2022-04-13T18:20:56.000Z (about 4 years ago)
• Default Branch: main
• Last Pushed: 2022-12-09T18:12:18.000Z (over 3 years ago)
• Last Synced: 2025-01-12T21:33:25.311Z (over 1 year ago)
• Topics: deep-learning, language-model, nlp, python, pytorch, reinforcement-learning
• Language: Python
• Homepage: https://sea-snell.github.io/CALM_LM_site/
• Size: 58.6 KB
• Stars: 34
• Watchers: 3
• Forks: 7
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# 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](https://sea-snell.github.io/CALM_LM_site/) | [arxiv](https://arxiv.org/abs/2204.10198)

## **setup**

1. create conda environment: `conda create --name CALM python=3.9.7`

2. activate conda environment: `conda activate CALM`

3. install requirements: `pip install -r requirements.txt`

4. install pytorch 1.9.0: `conda install pytorch==1.9.0 cudatoolkit=11.3 -c pytorch -c conda-forge`

5. `export PYTHONPATH="$PWD/offline_airdialogue"`

6. Download the processed data and model checkpoints [here](https://drive.google.com/drive/folders/1mnAGcgqyQC3ygILwwf-llxLf70nT9AT9?usp=sharing). 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 ` with `python -m torch.distributed.launch --nproc_per_node  --use_env `.

* **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 `



        * 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=`

* **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  --use_env language_quality_eval.py`