search

facebookresearch / synsin

View synthesis for the public.
http://www.robots.ox.ac.uk/~ow/synsin.html
Other
662 stars 94 forks source link

Discrepancy in sequence of operations inside ResNet block #25

Open VjayalakshmiK opened 4 years ago

VjayalakshmiK commented 4 years ago

Hi, I understand, from the paper and supplemental, that the ResNet block used in this work has been borrowed from that of BigGAN paper. The same has also been clearly illustrated in Figure 8(a)-(c) in the supplemental(image attached below).

However, the ResNet block definition, as done in blocks.py of model.layers (screenshot of code segment attached), does not seem to match the sequence of operations illustrated in the diagrams.

While in block.py, the order, for channel a, is normalization1-> activation1 -> conv1 -> normalization2 -> activation2 -> conv2 -> down/up/identity, in BigGAN as well in Figure 8 of your supplemental, down/up/identity comes before conv1 and after activation1 (not at the end, as has been coded). This is true of channel b as well.

Would it be possible to clarify which of the two is right? If the code is wrong, it needs to be rectified right away, given many have shown interest in your work :) ann1 ann2

oawiles commented 4 years ago

Yes you're right it's not the same. While it is inconsistent I'm not sure it's actually 'wrong' per se; more a different way of doing it. You could try the precise BlockGAN approach but I'm not sure it would give particularly different results.