How to check what is amount of padding used when we use padding = "same"?

1

Hello there,

I was wondering what is the amount of padding used when I put padding = “same” in conv2D!

I have an image.

import torch 

image = torch.randn((1, 3 ,224, 224))

conv2 = torch.nn.Conv2d(
    in_channels=3, 
    out_channels=5, 
    kernel_size=16,
    stride=1,
    padding='same'
)

conv2(image).shape  # 1, 5, 224, 224

But how to know what amount of padding has been used?
Since I have manually checked using padding = 4,5,6,7,8.
None of the padding gives 224 as output shape.

So that means is it using fractional padding?

Jay

2

Yes!

Not sure unfortunately. Afaik, it’s been handled internally. There may be some way to print it.

I think this is a nice hardcoded way to check?

kernel_size = conv2.kernel_size
if isinstance(kernel_size, int):
    padding = kernel_size // 2
else:
    padding = (kernel_size[0] // 2, kernel_size[1] // 2)

print(f"Padding used: {padding}")

output = conv2(image)
print(f"Output shape: {output.shape}")
3

Thanks @Soumya_Kundu for your reply!

With the code you provided, value of padding is coming as 8, but if we put 8 back, its nor preserving the image dimension.
image

Jay

4

Can anyone please help here?

5

You could check conv2._reversed_padding_repeated_twice and use F.pad before applying the convolution to get the same output, but keep in mind it’s an internal attribute and could change in the future.