Where is the Channel-wise fully-connected layer about context-encoder

[pathak22]

Hi ! The architecture I used in this repository is pretty simple. At the transition bottleneck, there are 4000 outputs for which if normal fully connected layer is used, it would result into 16M parameters which is usually fine (for e.g. AlexNet has more than that). This is described in Figure-9(a) in paper.

But if we were to use AlexNet until pool5 as generator then using normal FC would lead to about 100M parameters in single layer, hence channelwise-FC would be useful there. See Figure-9(b) in paper. For simple graphics inpainting task, we do not need to put AlexNet until pool5 architecture in generator. That is required for feature learning.