A possible bug in torch.nn.functional.conv1d/2d

liangbright · May 1, 2019, 2:43am

Pytorch 1.1.0, windows

import torch
import torch.nn as nn
import torch.nn.functional as nnF
x=torch.rand((16, 3, 28, 28))
a=nn.Conv2d(3,10,10,padding_mode='zeros')
y=nnF.conv2d(x, a.weight, padding_mode='zeros')

TypeError: conv2d() got an unexpected keyword argument ‘padding_mode’

Tony-Y · May 1, 2019, 8:00am

github.com

pytorch/pytorch/blob/142c973f4179e768164cd578951489e89021b29c/torch/nn/modules/conv.py#L330-L338


def forward(self, input):
    if self.padding_mode == 'circular':
        expanded_padding = ((self.padding[1] + 1) // 2, self.padding[1] // 2,
                            (self.padding[0] + 1) // 2, self.padding[0] // 2)
        return F.conv2d(F.pad(input, expanded_padding, mode='circular'),
                        self.weight, self.bias, self.stride,
                        _pair(0), self.dilation, self.groups)
    return F.conv2d(input, self.weight, self.bias, self.stride,
                    self.padding, self.dilation, self.groups)

F.conv2d does not have the optional argument padding_mode.

https://pytorch.org/docs/stable/nn.html#torch.nn.functional.conv2d

This document for F.conv2d is wrong.

ptrblck · May 1, 2019, 7:18pm

It looks like the circular padding is indeed applied before calling the functional conv API.
Could you open an issue so that the docs can be fixed?