Unlike keras, PyTorch inherently doesn’t have same padding options when defining convolutional layer. Does fast.ai have built-in function to support this functionality? If no, is there any clean way to implement same padding on convolutional layer?
Refer to @Atom-101’s and my reply for detailed suggestion
the solutions I found in PyTorch required a formula to calculate number of padding needed. A subclass has to be dedicatedly defined for same convolution layer.
I would like to see if anyone has a cleaner solution for that.
The output size for a conv operation is:
O = (W - F + 2P)/S + 1; where W is input image size, F is kernel size and S is stride
From this formula we can say:
P = ((S-1)W + F - 1)/2
P = (F - 1)/2; if S=1
In fact fastai’s conv_layer() function sets pad to (ks-1)/2 by default.
The only time this formula doesn’t work is when ks-1 is odd, meaning ks is even
(which is usually not case). In these cases you have to unevenly pad the two sides of an image. For example if you have to pad width by 1.5 pixels, you can pad the left of the image with 1 pixel and right with 2. You can use something like nn.ZeroPad2d() to do this.
The convention Tf/Keras follows in case of uneven pad is that it adds more pixels to the right and bottom than top and left.
@Atom-101 Thanks for the address!
I would like to summarize and supplement a few examples for those readers with the same question in mind.
1.If kernel_size is odd, you can get same padding from conv_layer() by default
It’s because of the default setting of conv_layer(). I extracted related source code from here for your reference:
if padding is None: padding = (ks-1)//2 if not transpose else 0
you can quickly test this out by the following experiments:
In [20]: from fastai.layers import conv_layer
In [21]: import torch
In [22]: t = torch.randn(1, 1, 5, 5)
In [23]: conv_3x3 = conv_layer(ni = 1, nf = 3, ks = 3)
In [24]: conv_3x3(t).shape
Out[24]: torch.Size([1, 3, 5, 5])
In [25]: conv_4x4 = conv_layer(ni = 1, nf = 3, ks = 4)
In [26]: conv_4x4(t).shape
Out[26]: torch.Size([1, 3, 4, 4])
As shown above, conv layer with kernel size = 4 fails to preserve input size (i.e. width and height)
2.if kernel size is even, you can still get same padding by conv_layer() augmented with nn.ZeroPad2d()
In this case, the padding is not evenly spread on 4 sides (i.e. top, bottom, left, right). you need to add extra padding by nn.ZeroPad2d():
In [27]: import torch.nn as nn
In [28]: extra_pad = nn.ZeroPad2d((0, 1, 0, 1)) # (left, right, top, bottom)
In [29]: conv_4x4(extra_pad(t)).shape
Out[29]: torch.Size([1, 3, 5, 5])
One more note is that nn.ZeroPad2d should come first, then conv_layer()