Loss function for adversarial example generation

[luan-xiaokun]

The loss you used for adversarial example generation seems to be the `myloss' function in evaluate.ipynb, but it does not match with the divergence and diversity score mentioned in the paper. Why did you use such a loss?