This repo includes the trained model (TOP_UWF_ema_model.pt / jitted version: TOP_UWF_ema_model_jit.pt) and an example for how to load the model and apply it to new images: MinimalModelLoadingExample_And_TestIO.ipynb.
The code was developed and run on Ubuntu 20.04.
You can use the requirements.txt to install the necessary packages. However, it was generated with pip freeze and might contain superfluous dependencies. We provide it to document the exact versions of all packages that were used. Alternatively, install pytorch according to the instructions at https://pytorch.org/get-started/locally/ and then run pip install timm==0.5.4 sklearn matplotlib tqdm pandas numpy notebook which should install all necessary dependencies. This should take a few minutes.
In the code, change PATH/TO/RAW/DATA/ to the path containing the unzipped raw data, as it was originally provided by Dr. Hiroki Masumoto, and PATH/TO/PREPROCESSED/DATA/ to the path where you wish to store the processed data. Unfortunately, we are unable to redistribute the data ourselves.
Then run TOP_preprocessing.py to preprocess the data and use TOP_Datasplit.ipynb to generate the train, validation and test sets. You should now be able to train a model with TOP_Training.ipynb.
A recent NVIDIA GPU is recommended for model training.