monodepth
monodepth copied to clipboard
mrharicot
ResNet-50
@mrharicot hi, thanks for your share. there are still some questions i cant understand.
-
the architecture design about resnet50 in the definition of 'resblock': in the standard resnet-50, the shortcut connection should located at the top layer when entering a resblock, just like below: def resblock(self,x,num_layers,num_blocks): out = x out = self.resconv(out, num_layers, 2) for i in range(num_blocks - 1): out = self.resconv(out, num_layers, 1) return out
is it right? or are there some other intention for the original design in your code?
Hi @7LFB, I think you are correct. My implementation of Resnet50 was quite last minute and has actually another bug. I will leave it as is for now to make sure people can reproduce the results from the publication. I however plan on having either a branch or a fork to fix all the bugs at some point in the near future.
Hi, are you sure this is a bug?
From my point of view, the shortcut connection is just x, so the line you were saying might be just the convolution of number of layers. The actual convolution which complete the job to resblock is inside def resconv. (the last line) return tf.nn.elu(conv3 + shortcut)
