DLDLM

This repository contains the implementation of the Discrete Latent Dialogue Language Model (DLDLM).

Repository structure

This repository is organised into four main directories:

• experiments/ contains the directories to host:
  • results of the experiments;
  • checkpoints generated during the experiments;
  • experiment configuration dumps;
  • experiment logs.
• notebooks/ contains the directories to host:
  • data exploration notebooks.
• resources/ contains:
  • directories to host the dialogue corpora used in the experiments, and the references to download them;
  • directory to host the YAML configuration files to run the experiments.
  • directory to host the pre-trained models, and the references to download them.
• src/ contains modules and scripts to:
  • run training and evaluation steps;
  • interact with the trained models;
  • load and preprocess corpora.

For further details, refer to the README.md within each directory.

Environment

To install all the required packages within an anaconda environment, run the following commands:

# Create anaconda environment (skip cudatoolkit option if you don't want to use the GPU)
conda create -n dldlm python=3.10 cudatoolkit=11.3
# Activate anaconda environment
conda activate dldlm
# Install packages
conda install pytorch=1.11.0 torchvision=0.12.0 -c pytorch
conda install -c conda-forge transformers=4.18.0
conda install -c conda-forge tensorboard=2.9.1 pandas scikit-learn matplotlib seaborn spacy jupyterlab
python -m spacy download en_core_web_sm
pip install spacytextblob
conda install -c plotly plotly
conda install -c conda-forge python-kaleido

To add the source code directory to the Python path, you can add this line to the file ~/.bashrc

export PYTHONPATH=$PYTHONPATH:/path/to/dldlm/src

Training

Run

There is a script to train or fine-tune the model, it expects to have ./src in the Python path and all data sets to be downloaded and placed in the ./resources/data/raw/ directory.

To train or fine-tune the model run:

python ./src/bin/train_dldlm.py --config_file_path ./resources/configs/path/to/config.yaml

To train or fine-tune the model in background run:

nohup python ./src/bin/train_dldlm.py --config_file_path ./resources/configs/path/to/config.yaml > experiment_"$(date '+%Y_%m_%d_%H_%M_%S')".out &

Monitor

To connect to a remote server and monitor the training process via Tensorboard connect via ssh to your machine using a tunnel

ssh  -L 16006:127.0.0.1:6006 user@adderess

Start the Tensorboard server on the remote machine

tensorboard --logdir ./expertiments/path/to/tensorboard/

Finally connect to http://127.0.0.1:16006 on your local machine

Evaluation

There is a script to run the final evaluation of the model, the requirements to run it are the same of the training script.

To run the evaluation in background execute:

nohup python ./src/bin/evaluate_static.py --config_file_path ./resources/configs/path/to/training/config.yaml > experiment_"$(date '+%Y_%m_%d_%H_%M_%S')".out &

Chatting

There is a script available to chat directly with any of the models, it can be run using the following command:

python ./src/bin/evaluate_interactive.py --config_file_path ./resources/configs/path/to/inference/config.yaml

Alternatively there is the chatbot_api sub-module designed for the re-use of the agent outside the repository. The API uses the base GPT-2 class from the Transformers library instead of the DLDLM extension we provide to promote re-use (note that it is not mandatory to install the other libraries in this case).

Here follows a usage example

from dldlm.chatbot_api.chatbot import DLDLMChatbot

# Define generate parameters (see HuggingFace Transformers documentation)
generate_kwargs = {'top_p': 1.0, 'top_k': 0, 'temperature': 0.7, 'do_sample': True}
# Create chatbot instance
chatbot = DLDLMChatbot(
  './resources/models/therapy_dldlm',
  None,
  max_context_len=256,
  max_response_len=256,
  generate_kwargs=generate_kwargs
)
# Define context (list of previous turns)
context = ['Hello, how are you?', 'Fine thanks, what about you?']
# Generate response
response = chatbot(context)