transIII_1x1_kxk does not behave as expected

[CoinCheung]

Hi,

I verified like this:

conv1 = nn.Conv2d(32, 64, 1, 1, 0, bias=True)
conv2 = nn.Conv2d(64, 128, 3, 1, 1, bias=True)
conv = nn.Conv2d(32, 128, 3, 1, 1, bias=True)

k, b = transIII_1x1_kxk(conv1.weight, conv1.bias, conv2.weight, conv2.bias, 1)
conv.weight.copy_(k)
conv.bias.copy_(b)
inten = torch.randn(2, 32, 224, 224)
out1 = conv2(conv1(inten))
out2 = conv(inten)
print((out1 - out2).abs().max())

And the output is 0.11, which is much too great. Have you noticed this ?

[DingXiaoH]

This is due to the padding.
In the paper, we explained this.
Notably, for a K×K conv that zero-pads the input, Eq. 8
does not hold because F (2) does not convolve on the result
of I ~ F (1) + REP(b(1)) (but an additional circle of zero
pixels). The solution is to either A) configure the first conv
with padding and the second without, or B) pad by b(1). An
efficient implementation of the latter is customizing the first
BN to 1) batch-normalize the input as usual, 2) calculate
b(1) (Eq. 6), 3) pad the batch-normalized result with b(1),
i.e., pad every channel j with a circle of b(1) j instead of 0.

In this repo, BNAndPadLayer is designed to solve this problem.

[CoinCheung]

Thanks for replying !!!