tstandley / xception-pytorch Goto Github PK

Hi,

Can you please add license to your code. Otherwise it cannot be used by others.

Thanks

Hi @tstandley,

Thanks for the great implementation. I tried to use the pretrained weights but it gives me shape mismatch error for each conv layer. Below is a sample output for one layer. My pytorch version is 0.4.

While copying the parameter named "block1.rep.0.conv1.weight", whose dimensions in the model are torch.Size([64, 1, 3, 3]) and whose dimensions in the checkpoint are torch.Size([64, 3, 3]).

hi~, I find that the forward function of SeparableConv2d class is different from Xception paper definition.
In Xception, the order of separable convolution operation should be 1x1 convolution first(pointwise convolution) and then depthwise convolution.
what do you think?

The model url "https://www.dropbox.com/s/1hplpzet9d7dv29/xception-c0a72b38.pth.tar?dl=1" doesn't work, would you update a new model weights?

hello, i can't get 'https://www.dropbox.com/s/1hplpzet9d7dv29/xception-c0a72b38.pth.tar?dl=1'.could you please give me your zip link?

In the paper, we see that in Depthwise Separable Convolution, first we pass the input through Conv1x1 and then through conv3x3. That is the order of the operation changes. But in the code in this repo and other repos, its done otherwise. Am I missing anything or my understanding is wrong?