Cookiecutter - ML Experiments in AWS, The Easy Way

Cookiecutter template for seriously scaling up hyperparameter tuning and ML experiments.

• GitHub repo: https://github.com/Yann21/cookiecutter-exp-launcher/
• Free software: BSD license
• Read Medium companion article

Features

• Select number of AWS instances on which to run the experiment.
• WanDB integration: Visualize results in real time with parallel plot.
• Plug and play: Change Python code to include your own model.

Quickstart

Install the latest Cookiecutter if you haven't installed it yet (this requires Cookiecutter 1.4.0 or higher):

pip install -U cookiecutter

Generate a Python package project:

cookiecutter https://github.com/Yann21/cookiecutter-exp-launcher.git

Then:

• Follow cookiecutter instructions and enter AWS and WandDB credentials.

$~ cd ml-exp/

Initialize AWS resources (need be done only once)
$~ ./cli.py init

Build and send the Docker application to AWS
$~ ./cli.py docker

Pull the trigger
$~ ./cli.py run 20

Room for improvement

There are countless ways in which you can improve the application. Here are some ideas:

• Use Bayesian optimization instead of grid search for faster convergence (WandB).
• Create your own EC2 instances instead of using the pricier Fargate.
• Initialize a spot ECS cluster.
• [...]

TODO

• Add e-mail notifier when the experiment is finished

Submit a Pull Request

I accept pull requests on this.