search

hytseng0509 / CrossDomainFewShot

Cross-Domain Few-Shot Classification via Learned Feature-Wise Transformation (ICLR 2020 spotlight)
323 stars 62 forks source link

The learned transformation layers are not used when testing? #14

Open chmxu opened 4 years ago

chmxu commented 4 years ago

Hi.

It seems that the scale and shift are omitted in all feature transformation layers when test. https://github.com/hytseng0509/CrossDomainFewShot/blob/master/methods/LFTNet.py#L189 https://github.com/hytseng0509/CrossDomainFewShot/blob/master/methods/backbone.py#L207. This means when testing, the proposed method is the same as the base model. Is this right?

hytseng0509 commented 3 years ago

Yes, the feature-wise transformation layers are used only during the training phase to improve the model generalization.