proceduralia / pytorch-conv2_1d Goto Github PK

Hello,

It looks like you're missing a ReLU between conv2d and conv1d in your module's forward declaration.

Section 3.5, 2nd paragraph of the paper.

Thanks,
Alex

def forward(self, x):
    b, c, t, d1, d2 = x.size()
    x = x.view(b*t, c, d1, d2)
    x = self.conv2d(x)
    dr1, dr2 = x.size(2), x.size(3)
    x = x.view(b*dr1*dr2, x.size(1), t)
    x = self.conv1d(x)
    x = x.view(b, self.out_channels, -1, dr1, dr2)
    return x

here should not we permute the dim first?
b, c, t, d1, d2 = x.size()
+++x=x.permute(0,2,1,3,4)
x = x.view(b*t, c, d1, d2)
otherwise,the meaning of x is not the conv2d needed like:[batchsize,channel,height,width]