bug about fft-conv-pytorch HOT 4 CLOSED

nn.conv2d:
def init(
self,
in_channels,
out_channels,
kernel_size,
stride,#<==
padding,#<==
dilation,
groups,
bias,
padding_mode
)

ffconv2d:
def init(
self,
in_channels,
out_channels,
kernel_size,
padding,#<==
stride,#<==
dilation,
groups
)

from fft-conv-pytorch.

Hi -- I'm not sure this is a bug? The weights/biases for nn.Conv2d and FFTConv2d are randomly initialized, so in your example they will be different for the two layers. Can you manually set the weights/biases to be equal, and see if the outputs are still different?

from fft-conv-pytorch.

Not a bug, the padding and the stride switch places

from fft-conv-pytorch.

Understood. I'll switch the positions of those arguments, to match the form of f.conv2d.

By the way, you can format code in Markdown using backticks: https://www.freecodecamp.org/news/how-to-format-code-in-markdown/
Makes your code snippets much easier to read.

from fft-conv-pytorch.