xudonglinthu / context-gated-convolution Goto Github PK

Hello, I am getting the following error when I run your content gate conv. How do I fix this error?
return torch.matmul(out, x_un).view(b, self.oc, int(np.sqrt(l)), int(np.sqrt(l)))
RuntimeError: shape '[1, 64, 69, 69]' is invalid for input of size 306432

Hello, I would like to ask you what is the purpose of using a pooling layer to make the feature map of h*w into h'*w'? AdaptiveAvgPool2d((ws,ws)) in the code you posted, why output_size=kernel_size?

Nice work!
I wonna use context-gated-convolution in action recgnition. I noticed that you use torch.nn.Unfold in layer.py, and it needs a 4-D input(batch x channel x height x width). But I don't know how to do to extend context-gated-convolution to a 5-D input (batch x channel x frams x height x width) . Will you release your code for action recgnition in the paper?
Thanks!

Hello, I find your job very creative and utilize it on my network. However, the GPU consumption enlarges four times as the former network when training. Have you ever encountered this problem and how can I address it? Thank you very much for your reply!

in layers.py
num_lat -> self.num_lat

At line 19 in context-gated-convolution/image-classification/models/layers.py,

if kernel_size == 1 or True:
    self.ind = True

self.ind is always True and the layer works the same as traditional convolution.

Is or True redundant?