zootopia3 0.1

Fitting simple torch models

Fitting PyTorch models

This repository follows a workshop to set up a Python package, build some neural networks with torch, and publish the models on HuggingFace.

There are two datasets that we simulate:

1. Colored shapes (circle, rectangle, triangle, diamond) in a pixellated image
2. One-dimensional sinusoids

Alongside these datasets, we fit the following model objectives:

1. (Classification) Predict the shape and color in the image
2. (Regression) Predict the next time steps of the sine function

Usage

Learn how to train neural networks from scratch.

1. Run answers/simulate-exercise.ipynb to get data.
2. Fill in the # TO DO parts in examples/modeling-exercise-*.ipynb.
3. Compare to solutions in answers/modeling-exercise-*.ipynb.
4. You can explore different parameters on big models with modeling.py.
   • Write a shell script that invokes modeling.py and pass to slurm.

The package defined under src/ provides:

• A class Shape that instantiates an image with 1 colored shape
• A function simulate_shapes() to make many images for an image classifier
• A model class MyCNN to fit a standard architecture

Caution: you may need GPU resources if your models or data are large.

Requirements

• Python 3.10+

Install

If you want to install the package only from the internet:

pip install zootopia3

If you want to set up an isolated environment and build locally:

python -m venv path-to/your-environment
source path-to/your-environment/bin/activate
pip install -e .

You run the pip command within this repo.

Test

You can run the test scripts in tests/ with the following:

python -m pytest