Dataset2Vec

Introduction

This package aims to implement the approach proposed in Dataset2Vec: Learning Dataset Meta-Features by Jomaa et al. This package makes the training Dataset2Vec dataset encoder much more approachable by providing an API that is compatible with pytorch-lightning's trainer API. The output logs including tensorboard and checkpoints are stored in lightning_logs or in default_root_dir from pytroch_lightning.Trainer if specified.

Installation

To install the package run the following command (you need Python 3.9 or higher):

pip install -r requirements.txt

Usage

Here is a simple example of the usage of the package:

from pathlib import Path

from pytorch_lightning import Trainer

from dataset2vec import (
    Dataset2Vec,
    Dataset2VecLoader,
    RepeatableDataset2VecLoader,
)

train_loader = Dataset2VecLoader(Path("data/train"))  # Path with .csv files
val_loader = RepeatableDataset2VecLoader(
    Path("data/val")
)  # Path with .csv files

model = Dataset2Vec()

trainer = Trainer(
    max_epochs=2, log_every_n_steps=1, default_root_dir="output_logs"
)  # output of the training will be stored in output_logs

trainer.fit(model, train_loader, val_loader)

Development

Here are the snippets useful for the development of the package:

• ./scripts/check_code.sh - runs code quality checking using black, flake8, isort and mypy.
• pytest - runs all unit tests
• cd docs && make html - generates documentation
• python -m build - build the package
• twine upload dist/* - uploads the package to PyPI