Skip to content

Latest commit

 

History

History
112 lines (78 loc) · 6.31 KB

README.md

File metadata and controls

112 lines (78 loc) · 6.31 KB
Raw

SAINT: Improved Neural Networks for Tabular Data via Row Attention and Contrastive Pretraining

saint architecture

Paper Reference: https://arxiv.org/abs/2106.01342

NB: This implementation uses Pytorch-lightning for setting up experiments and Hydra for configuration. For an earlier release of this code which does not use hydra, check the branch saint-orig

Motivation

We decided to create an implementation of saint that can work with any tabular dataset, not jsut those mentioned in the paper. This implementation can be run with any tabular data for binary or multiclass classification. Regression is not supported.

Code Structure

Major modules implemented in the code

  • Saint Transformer
  • Saint Intersample Transformer
  • Embeddings for tabular data
  • Mixup
  • CutMix
  • Contrastive Loss
  • Denoising Loss

For easy configuration, we decided to organize the code in a structured way

  1. The base config file can be found the the configs directory. This contains most hyperparameters for training the models.
  2. Configurations for experiments are split into supervised and self-supervised configs. The supervised config file is to be used as default when running predictive tasks while the self-supervised config file should be selected in case of self-supervised pre-training tasks. The files can be edited to suit your needs. In addition, there is a predict config file which we setup to make predictions on a test set (e.g. for kaggle)
  3. Another config directory is dedicated to house all datasets configurations. This is the data sub-directory inside the configs directory. It includes hyperparameters like train, validation and test data paths and other data statistics. Samples of configs for supervised (indicated with names ending in "_sup") and self-supervised training (indicated with names ending in "_ssl") are provided for bank dataset. They can be replicated for other custom datasets as well.

Dataset

The datasets should live in the data directory. It is necessary to provide the absolute paths to the data folder in the data configs. Also, the datasets have to be pre-processed before running experiments. These are recommendations from the paper e.g data transforms using z-transform. Other recommendations are design decisions made by us.

Process your dataset in the following format

  • Add a column named 'cls' to your dataset. The 'cls' column should be the first column as mentioned in paper
  • Apply z-transform to numerical columns
  • Label encode categorical columns
  • It is required that categorical columns are separated from numerical columns. In particular, we track the total categorical columns in the dataframe by the number of columns. Therefore, you need split the data into categorical columns and numerical columns, compute some statistics as explained below, then merge them to form a new dataframe which can be understood by the model. Note that all categorical columns appear before the numerical columns in the data. The cls column is expected the be the first column and it is counted as a categorical column.
  • Calculate the number of categorical columns (including 'cls' column) and numerical columns. Include these statistics under data_stats in the data configs of your particular dataset.
  • Also, you will need to provide the number of categories in each categorical column in the data config. These are required to build proper embeddings for the model. It should also be provided as an array in the data.data_stats.cats parameter or hard coded in the data config. Note that the cls column has 1 category and is always indicated as 1 in the array.
  • Save your processed files as train, val and test csvs in data folder.

A sample function named preprocess.py is included under src > dataset.py that explains the preprocessing strategy. You may need to modify this function depending on the dataset. Also, tutorial notebooks are provided in notebooks folder to showcase how to preprocess custom datasets and run experiments. Look at Bank_Dataset.ipynb

How to set it up

Clone the repository

git clone https://github.com/ogunlao/saint.git

Setup a new environment

  1. Activate your virtual environment. It is advisable to use a virtual environment to setup this code.
  2. Install dependencies using the requirements.txt file provided
pip3 install -r requirements.txt
  1. Update the config.yaml file with your hyperparameters. Alternatively, you can provide your settings on the command-line while running experiments. A good knowledge of hydra might be required.

Run python main.py with command-line arguments or with your edited config file

Examples

  1. To train saint-intersample (saint-i) model in self-supervised mode using bank dataset, run;
python main.py experiment=self-supervised \
    experiment.model=saint_i \
    data=bank_ssl \
    data.data_folder=/content/saint/data
  1. To train saint model in supervised mode using bank dataset, run;
python main.py experiment=supervised \
    experiment.model=saint \
    data=bank_sup \
    data.data_folder=/content/saint/data
  1. To make prediction using saint model in supervised mode using bank dataset, run;
!python saint/predict.py experiment=predict \
    experiment.model=saint \
    experiment.pretrained_checkpoint=["PATH_TO_SAVED_CKPT"] \
    experiment.pred_sav_path=["PATH_TO_SAVE_PREDICTION.csv"] \
    data=bank_sup \
    data.data_folder=/content/saint/data

You may need to run some hyperparameter search to determine the best model for your task. Hydra provides this functionality out of the box with multirun.

Contributors

(names in alphabetical order)