tezer/canard_test

Canard_test

This is a simple command-line utility for anaphora resolution in short dialogs. There is a trainer (train_t5.py) and a model runner (run_model.py). The file constants.pycontains the most important settings for training and for the task.

I used T5 t5-small for this task as the most appropriate type of transformer.

The main libraries used in the project are: transformers, pandas, torch and pytorch_lightning. The latter makes the PyTorch-based code more readable.

Here is what the code does for training:

1. loads the data to DataFrame
2. defines a callback that tracks validation loss, and best checkpoint saver to make sure I have the best trained model
3. finally, packs everything into the Trainer and fires it off

run_model.py is much simpler:

1. loads the saved model (a checkpoint file)
2. gets rewritten question with resolved anaphora from the model for a user input
3. and writes them to the screen

Installation

clone the project with `git clone https://github.com/tezer/canard_test

Usage

Requires python 3.8 cd canard_test

For model training run:

poetry run python canard/train_t5.py for interactive mode

or

poetry run python canard/train_t5.py --input_file=/path/to/your/data/folder

where

Options:
  --input_file    Specify the input folder with dev.jsom, train.json and test.json files.
  --help          Show this message and exit.

To run the model:

poetry run python canard/run_model.py for interactive mode

or

poetry run python run_model.py --model_file=/path/to/your/best-checkpoint.ckpt

where

Options:
  --model_file FILENAME   Enter the path to the file with the trained model.
  --help                  Show this message and exit.