Discriminator feature matching loss

[leicaand]

The "Weak Feature Matching Loss" is different (both here and in the original repo), from the one mentioned in the paper (the features from all the layers are used, instead of the last few layers)

[a312863063]

From my training experience, it is better to add feature matching loss to all layer features of D. This operation is conducive to achieving a better training balance.

[gitlabspy]

From my training experience, it is better to add feature matching loss to all layer features of D. This operation is conducive to achieving a better training balance.

I noticed that getIntermFeat in discriminator is set to False as default which makes D returns only last layer output. If all layer features of D should be used for wfm, should getIntermFeat be set to True?