Unused feature in forward pass? about gscnn

[tovacinni]

Thanks for your interest in our work!

You are right, s1 is not needed- that's likely a variable we forgot to delete. m1 comes from the first layer in the resnet, so it follows the paper- s1 is essentially a duplicate variable. Will close this issue when I delete this.

Let me double check, but I believe you are right.

Following the paper,

[tovacinni]

3648b86

[rajexp]

@tovacinni When I look at the paper there the Figure 2. All shape stream inputs from regular stream seem to be passing through conv1x1.

Going by that s1 should have been used in the first residual block res1 contrary to what is presented in the code where m1 being direct feature (conv3x3) of the regular stream is used.

Could you please look into this. Any possible explanation is appreciated.