Replicate MURA baseline and playing around

Mura competition https://stanfordmlgroup.github.io/competitions/mura/

Please Use Pull-Request If You Want To Update Master Branch, And Add sx5640 As Reviewer

How to run the script

1. Setup your preferred environment, i.e. virtualenv, conda, docker
2. Install packages: pip install -r requirements.txt
3. Download MURA dataset into /dataset/ directory. Link to MURA: https://stanfordmlgroup.github.io/competitions/mura/
4. To run a model on MURA dataset, run python {path-to-model-python-file} train. Add -h flag for all options.
5. To show a visualization of model attention or activation of a layer, run python visualize.py. Add -h flag for all options.

Note:

• In order to run tensorflow in cpu, install tensorflow instead of tensorflow-gpu in requirements.txt
• For MacOS user, importing dataset might throw exception, as explained by: https://matplotlib.org/faq/osx_framework.html If you use virtualenv, one solution pointed out by Kevin is to add the following code to the top of your dataset.py:

  import matplotlib as mpl
  mpl.use('TkAgg')

• The main script does not work with keras-2.2.0 due to a known bug in keras, but will work with keras-2.2.2.