This is a re-implementatino of https://github.com/hszhao/SAN/blob/master/model/san.py SAN model. We translated this model to tensorflow and used it to classify bone x-ray photos, achieving a 78.9% accuracy.
Command "python model/train.py" to run on sample_dataset of 2000 images (training + testing combined)
- Ask permission from https://stanfordmlgroup.github.io/competitions/mura/
- !curl https://cs.stanford.edu/group/mlgroup/MURA-v1.1.zip --output m.zip
- !unzip m.zip
- !mkdir whole_dataset
- !mkdir whole_dataset/train
- !mkdir whole_dataset/test
- !python mura_dataset_reorganize.py
- Finally, change directory in model/train.py from sample_dataset to whole_dataset
- Tune hyper parameters in config.py
- You can tune kernal size and layer size in model/train.py
- You can tune channels in model/san.py