cnn not running correctly with trained models

[paullintilhac]

- python3 run_defense.py -m cnn --n_epoch=2 -d retrain 

this line in travis.yml seems to be failing. should be fixed with uttam's update to load cnn models properly. but Paul is still getting ValueError: Cannot create Layer with a non-positive input_shape dimension. input_shape=(None, 64, -3), self.name=None

[uttam-rao]

This is because based on the default we were trying to do the retrain defense for rd=10 (without rev) which doesn't exist in the nn_models folder and also is too small.

For running the retrain defense on the cnn the minimum rd we can use is 30. Anything below that and the input dimensions in the build_cnn_rd function will become negative which will throw an error (I remember seeing a comment about this in an older version of the code too).

I've changed the default rd to 100 in the data utils file and this passes travis with no issues.