This is a framework to gain insight into pytorch deep neural networks for visual recognition tasks.

The README.md contains only a very brief description - for details, visit the

--> full documentation and apidoc <--

The project is split into the following parts, with the linked jupyter notebooks explaining them:

• [uncertainty](notebooks/Monte-Carlo Uncertainty.ipynb): confidence of predictions
• visualization of network properties.
  • base building blocks in a consistent and flexible interface
  • [compound](notebooks/Compound Visualization.ipynb) methods implementing 'standard' visualization methods

To follow the examples interactively, clone the repository and run poetry install. Then start jupyter with poetry run jupyter notebook.

• uncertainties
  • predictive entropy (total uncertainty).
  • mutual information (model uncertainty).
• building blocks
• visualization
  • pixel activation optimization
  • (guided) backpropagation
  • (guided) gradCAM
  • occlusion

From package index: pip install "midnite" or from source:

git clone https://gitlab.com/luminovo/public/midnite.git
cd midnite
poetry build
pip install dist/midnite-*.whl

We value clean interfaces and well-tested code. If you want to contribute, usually it's best to open an issue first.

We use poetry to manage dependencies.

Please make sure to have the pre-commit hooks installed. Install pre-commit and then run pre-commit install to register the hooks with git.

We use make to streamline our development workflow. Run make help to see all available commands.

$ make help 
help                 Show this help message
check                Run all static checks (like pre-commit hooks)
docs                 Build all docs 

This project is under the MIT license.

Scientific sources: see reference doc page.

This project was developed as a student IDP at luminovo.ai

Code contributors:

• Fabian Huch
• Christina Aigner