get_fda_loss

[qilong-zhang]

Hi @BardOfCodes,

Recently, I am reading your paper "FDA: Feature Disruptive Attack”, which is really interesting work. But I am a little confused about your implementation, e.g., get_fda_loss(self, opt_operations): Why do you assign batch_size as layer.shape[0] // 2? I mean why mean across channels C(h,w) is not equal to tf.reduce_mean(layer[:], -1)? Could you help me, thanks!