I’m trying to implement an Image classification model, but I’m having trouble understanding how to go from a lower output shape to a higher output shape. Below is a picture:
How does the author of this paper go from Maxpool2d_2 to conv2d_3? The only way that I can imagine this happening is if you use nn.ConvTranspose2d(), however this is not mentioned in the paper. Is there an alternative way of doing this step?
The paper: https://www.sciencedirect.com/science/article/pii/S0960077920308870#sec0002
It’s unclear as the authors didn’t share enough information, but they might have used a small stride and large padding.
Here is an example:
conv = nn.Conv2d(128, 64, 3, 1, padding=7)
x = torch.randn(1, 128, 9, 9)
out = conv(x)
print(out.shape)
# torch.Size([1, 64, 21, 21])
print(torch.tensor([p.nelement() for p in conv.parameters()]).sum())
# tensor(73792)
As you can see it’s just a guess and I didn’t consider any other dilation argument, which might also be different. The number of parameters also doesn’t match, so take this code with a grain of salt.
