MET - PyTorch

This repo reproduces the MET (Masked Encoding for Tabular Data) framework for self-supervised learning with tabular data.

Authors: Kushal Majmundar, Sachin Goyal, Praneeth Netrapalli, Prateek Jain

Reference: Kushal Majmundar, Sachin Goyal, Praneeth Netrapalli, Prateek Jain, "MET: Masked Encoding for Tabular Data," Neural Information Processing Systems (NeurIPS), 2022.

Original paper: https://table-representation-learning.github.io/assets/papers/met_masked_encoding_for_tabula.pdf

Original repo: https://github.com/google-research/met

Install

Clone this repository, create a new Conda environment and

git clone https://github.com/chris-santiago/met.git
conda env create -f environment.yml
cd met
pip install -e .

Use

Prerequisites

Hydra

This project uses Hydra for managing configuration CLI arguments. See met/conf for full configuration details.

Task

This project uses Task as a task runner. Though the underlying Python commands can be executed without it, we recommend installing Task for ease of use. Details located in Taskfile.yml.

Current commands

> task -l
task: Available tasks for this project:
* check-config:       Check Hydra configuration
* compare:            Compare using linear baselines
* train:              Train a model
* wandb:              Login to Weights & Biases

Example: Train model and for adult-income dataset experiment

The -- forwards CLI arguments to Hydra.

task train -- experiment=income

PDM

This project was built using this cookiecutter and is setup to use PDM for dependency management, though it's not required for package installation.

Weights and Biases

This project is set up to log experiment results with Weights and Biases. It expects an API key within a .env file in the root directory:

WANDB_KEY=<my-super-secret-key>

Users can configure different logger(s) within the conf/trainer/default.yaml file.