cv-boilerplate

Open source boilerplate for computer vision research. We solve small problems related to project management, experiment management, code styling, and model training so you can get back to research!

Installation

Standard

git clone https://github.com/pennpolygons/cv-boilerplate.git
cd cv-boilerplate
make init

Anaconda

git clone https://github.com/pennpolygons/cv-boilerplate.git
cd cv-boilerplate
make init-conda

Demos

By default training uses: /research/configs/default.yaml

Training outputs:

• Visdom (in browser): localhost:8000
• Output log file, images, data files, hydra logs, Visdom logs: /outputs

Example: Regression

Train using default config

python research/train.py example=regression

Train using Hydra to modify default config at runtime

python research/train.py mode.train.max_epochs=1 example=regression

Example: Classification

Train using default config

python research/train.py example=classification

Train using Hydra to modify default config at runtime

python research/train.py mode.train.max_epochs=1 example=classification

Building Blocks & Choices

This project is a collection of principled choices to make CV research easier. We prioritize using tools that minimize researcher engineering work, help keep the codebase uncluttered, and make research easy to distribute and reproduce.

• Python Formatting: Black
• Configuration Management: Hydra
• Machine Learning Framework: PyTorch
• Training / Evaluation Managment: PyTorch Ignite
• Visualization: Visdom
• Typed Code: Types!
• Version Control: Git