Question about the parameter of upsampling in FCN32s
super233 opened this issue · 0 comments
super233 commented
According to the formula of ConvTranspose2d: output = (input - 1) * stride - 2 * padding + kernel_size. To realize 32 times upsampling, the parameter should be (kernel_size=32, stride=32, padding=0).
But the parameter in code is self.upscore = nn.ConvTranspose2d(num_classes, num_classes, kernel_size=64, stride=32, bias=False)
.
Could you please tell me why it's this? Thank you : )