About The Project

This is a template for a Kaggle project, adapted to the Feedback 3 Prize competition.

The goal is for a Kaggle competitor to be able to:

Develop their solution in an IDE (PyCharm, VSCode, etc) rather than in a Jupyter-style notebook
Easily train a model in different environments (Kaggle, Google Colab, Jarvislabs, etc)
Submit a solution without copy-pasting code to an inference notebook

Usage Guide

The project has a single entry point, which is the script run_solution. It can be used to run all three typical Kaggle workflows: train a model, create a submission and make predictions for ensembling.

Local Environment Setup

Create an IDE project from the repository root
Create a virtual environment for the project (using either virtualenv, conda, or a similar tool)
Install the project in development mode: cd src && pip install -e .

Train a Model

In the training environment, install the project: cd src && pip install .
Run the script: run_solution env=[ENVIRONMENT_NAME] solution=[SOLUTION_NAME] solution.do_train=true
Export the project to a wheel and export it to a Kaggle dataset

Example Notebook

Create a Submission

In a Kaggle notebook, install the project from a dataset: cd ../input/[DATASET NAME] && pip install .
Run the script: run_solution wandb.enabled=false env=[ENVIRONMENT_NAME] solution=[SOLUTION_NAME] solution.do_create_submission=true

Example Notebook

Configuration

Overview

The configuration of the run_solution script is done with the hydra-core package, which allows specifying any attribute via either the config yaml files or the CLI. The top-level configs are env and solution. Their respective yaml's are located in:

.
└── src                   
    └── kaggle_ell
        └── config
            ├── env
            │    ├── colab.yaml
            │    ├── kaggle.yaml
            │    ├── local.yaml
            │    └── [YOUR ENV NAME].yaml
            └── solution 
                ├── dummy_solution.yaml
                └── [YOUR SOLUTION NAME].yaml

Setting Configuration Parameters

Toggling between different environments and solutions is as simple as calling run_solution env=[ENV NAME] solution=[SOLUTION NAME].

It's also possible to override specific parameter values from the command-line (or notebook). Example:

run_solution \
env=kaggle \
solution=transformer_finetune \
solution.do_train=true \
solution.args.train.epochs=3 \
solution.args.train.trn_fold=[0,1,2,3] \
solution.args.model.backbone=microsoft/deberta-v3-base

mrcabbage972/kaggle-ell-public