ProtoTrees

Refactored version of https://github.com/M-Nauta/ProtoTree and parts of https://github.com/cfchen-duke/ProtoPNet to make them more modular and easier to use. This will probably be turned into a Python package and moved to a new repository.

Setup

For all datasets

1. Create a Python >=3.11 environment.
2. Install requirements from requirements.txt (e.g. pip install -r requirements.txt).
3. Install Graphviz. With the current code you need to be able to call dot from the terminal.
4. Install project code in editable mode using pip install -e .
5. You can train the tree model and see its performance on the test set with python src/run_model.py --model_type protopnet (or --model_type prototree).

   NOTE: src/util/args.py has a list of all args that can be used to configure the run.

For CUB dataset

1. Install requirements from datasources/requirements-download.txt.
2. Run python datasources/cub_download.py.
3. Run python datasources/cub_preprocess.py.
4. (Optional, but recommended) Download a ResNet50 pretrained on iNaturalist2017 (filename on Google Drive: BBN.iNaturalist2017.res50.180epoch.best_model.pth) and place it in the folder src/features/state_dicts.

For development

Currently, all these steps are only done manually on a development machine. We should set up a pipeline that does these things automatically and reproducibly.

1. Install requirements from requirements-dev.txt.
2. You can lint the code with black src tests.
3. You can check types with MYPYPATH=src mypy src tests --explicit-package-bases --check-untyped-defs. Note that it will be helpful to run mypy --install-types beforehand.
4. You can run tests with the command pytest.