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.

Run on sample_dataset

Command "python model/train.py" to run on sample_dataset of 2000 images (training + testing combined)

Run on whole_dataset

  1. Ask permission from https://stanfordmlgroup.github.io/competitions/mura/
  2. !curl https://cs.stanford.edu/group/mlgroup/MURA-v1.1.zip --output m.zip
  3. !unzip m.zip
  4. !mkdir whole_dataset
  5. !mkdir whole_dataset/train
  6. !mkdir whole_dataset/test
  7. !python mura_dataset_reorganize.py
  8. Finally, change directory in model/train.py from sample_dataset to whole_dataset

How to tune

  1. Tune hyper parameters in config.py
  2. You can tune kernal size and layer size in model/train.py
  3. You can tune channels in model/san.py