AvgPool2d's padding working incorrectly

[kwmaeng91]

For some reason, AvgPool2d's padding works incorrectly.
Below is the minimal example to reproduce:

crypten.init()
x = torch.rand([16, 3, 128, 128])
x_enc = crypten.cryptensor(x, src=0)
model = nn.AvgPool2d(kernel_size=3, stride=2, padding=1)
print(model(x).shape)
model_enc = crypten.nn.from_pytorch(model, dummy_input=(x))
model_enc.encrypt()
print(model_enc(x_enc).shape)

The first print prints (16, 3, 64, 64). However, the second print (with Crypten) produces (18, 3, 64, 63). It is padding the first (batch) dimension instead of the last.

I tracked it down a little bit, and found that in _ConstantPad function, input[1] is being given as (0, 0, 1, 1, 0, 0, 1, 1), where it should be (1, 1, 1, 1). Not sure why.