Description

This repository contains a full Pytorch Lightning pipeline with the main goal of reusing it as a template for Kaggle competitions. It is particularized to the FTTransformer model from Gorishniy, Rubachev, Khrulkov, & Babenko (2021, Nov) Revisiting Deep Learning Models for Tabular Data, but any Pytorch model can be used.

The code was tested using the data from https://www.kaggle.com/c/tabular-playground-series-dec-2021.

---

Folder structure for future projects

project name    
│
└───data
│   │   train.csv
│   │   test.csv
│   │   sample_submission.csv
│   
└───src
│   │   main.py
│   │   datasets.py
│   │   dataprep.py
│   │   model.py
│   
└───simulations
    └───simulation_name
        |   config.yml
        │   oofs.csv
        │   hpopt.pkl
        |   submission.csv
        └───checkpoints
            │   fold=0_epoch=[]_metric=[]_val_metric=[].ckpt
            │   fold=1_epoch=[]_metric=[]_val_metric=[].ckpt
            |   ...

---

Usage

This Pytorch Lightning pipeline can used for any Machine Learning problem with the following modifications:

• create_folds function inside the dataprep.py module
• prep_data function within the dataprep.py module
• Datasets within the datasets.py module
• prepare_data method within the datasets.py module
• Loss function within the model.py module
• Metric within the model.py and the main.py module
• The Pytorch model
• The way that predictions are made