Difference between conv2d and conv2dtranspose with kernel size 1

Coca Cola35 0
Viewed 23

I understand that conv2d is used for downsampling and conv2dtranspose is the opposite (upsampling). However, assuming we are not using stride or padding here. Is there a difference between the two?

Downsampling means reducing the size of input dimension. for example If you have an input of (Batch Size = 5, Channel = 3, Height = 8, Width = 8), if you reduce the height and width using maxpooling (stride=2 kernel_size=2) the output becomes (Batch Size = 5, Channel = 3, Height = 4, Width = 4). That's downsampling, the opposite is upsampling (Increasing the Height and Width dimension)

for example:

classifier1 = torch.nn.Conv2d(in_channels=10, out_channels=5, kernel_size=1)
classifier2 = torch.nn.Conv2dTranspose(in_channels=10, out_channels=5, kernel_size=1)
deep-learning neural-network conv-neural-network pytorch
1 Answers

Operation wise, no difference. ConvTranspose2d() inserts stride - 1 zeros inbetween all rows and columns, adds kernel size - padding - 1 padding zeros, then does exactly the same stuff as Conv2d(). Default arguments result in no changes.

Though, if you actually run them back to back like this on the same input, the results will vary unless you explicitly equalize the inital weights, of course.

dx2-660
Related