Travis FSLeyes installs failing

For the last 2 days, all Travis builds with FSLeyes install are failing on test_train_network.py.

The error comes from Keras and happened in train_network.py where loading the trained model.

Error ADS train_network.py

test/test_train_network.py:135: 
AxonDeepSeg/train_network.py:289: in train_model
    str(path_model) + "/model.hdf5", custom_objects=custom_objects
../../../miniconda/envs/ads_venv/lib/python3.7/site-packages/keras/engine/saving.py:419: in load_model
    model = _deserialize_model(f, custom_objects, compile)

Error Keras

        model_config = f['model_config']
        if model_config is None:
            raise ValueError('No model found in config.')
>       model_config = json.loads(model_config.decode('utf-8'))
E       AttributeError: 'str' object has no attribute 'decode'
../../../miniconda/envs/ads_venv/lib/python3.7/site-packages/keras/engine/saving.py:224: AttributeError

The same test with Basic Install is passing.

This may be linked with the issue we had with h5py version in PR #382.

The “basic install” installs h5py=2.10.0. The “FSLeyes install” with conda-forge seems to install h5py=3.1.0.

The last build with FSLeyes that passed was 5 days ago, and FSLeyes was installing h5py=2.10.0.

[mathieuboudreau]

Fair point re: h5py!

The problem with doing pip install -e . after installing FSLeyes is that we might downgrade other packages are needed for FSLeyes too, breaking that tool now. Maybe just redownloading the proper version h5py would be more prudent.

[mathieuboudreau]

Hmmm. I know that installing FSLeyes using conda overrides some of the requirement versions of our original ADS install sometimes. Could you try pip installing ADS, installing FSLeyes, then pip installing ADS again to see if that it resolves? If so, we could do a comparison of the pip freeze output before and after FSLeyes being installed.