Using a temporal-difference variational autoencoder (Gregor et al, 2019) to predict abnormal cell division phenotypes in the mitocheck dataset (paper, data, metadata) (implented in pytorch).
All mitocheck movies should be downloaded and preprocessed using this github repo.
Then, combine all movies into a single pickle file using convert_videos.py.
This script expects that all preprocessesed movies are saved to their orginal output location and have not been moved.
main_train.py is used to train the model.
All hyperparameters except input_size and processed_x_size can be adjusted as desired.
input_size is the dimensions (height * width) of the preprocessed movies and should be changed only if a different compression factor was used during preprocessing.
Additionally time_constant_max and time_jump_options should be changed if a different number of frames was specified during preprocessing.
Training the model will generate two outputs.
The loss for each iteration is saved in loginfo.txt and the model is saved every 10 epochs to the output folder.