Performing Convolution (NOT cross-correlation) in pytorch

问题

I have a network (https://github.com/TheAbhiKumar/tensorflow-value-iteration-networks) that I am trying to implement in pytorch (I'm very new to pytorch, however, not at all new to machine learning).

In short, I cannot seem to figure out how to implement "pure" convolution in pytorch. In tensorflow it could be accomplished like this:

def conv2d_flipkernel(x, k, name=None):
    return tf.nn.conv2d(x, flipkernel(k), name=name,
                        strides=(1, 1, 1, 1), padding='SAME')

With the flipkernel function being:

def flipkernel(kern):
      return kern[(slice(None, None, -1),) * 2 + (slice(None), slice(None))]

How could something similar be done in pytorch?

回答1:

TLDR Use the convolution from the functional toolbox, torch.nn.fuctional.conv2d, not torch.nn.conv2d, and flip your filter around the vertical and horizontal axis.

torch.nn.conv2d is a convolutional layer for a network. Because weights are learned, it does not matter if it is implemented using cross-correlation, because the network will simply learn a mirrored version of the kernel (Thanks @etarion for this clarification).

torch.nn.fuctional.conv2d performs convolution with the inputs and weights provided as arguments, similar to the tensorflow function in your example. I wrote a simple test to determine whether, like the tensorflow function, it is actually performing cross-correlation and it is necessary to flip the filter for correct convolutional results.

import torch
import torch.nn.functional as F
import torch.autograd as autograd
import numpy as np

#A vertical edge detection filter. 
#Because this filter is not symmetric, for correct convolution the filter must be flipped before element-wise multiplication
filters = autograd.Variable(torch.FloatTensor([[[[-1, 1]]]]))

#A test image of a square
inputs = autograd.Variable(torch.FloatTensor([[[[0,0,0,0,0,0,0], [0, 0, 1, 1, 1, 0, 0], 
                                             [0, 0, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 0, 0],
                                            [0,0,0,0,0,0,0]]]]))
print(F.conv2d(inputs, filters))

This outputs

Variable containing:
(0 ,0 ,.,.) = 
  0  0  0  0  0  0
  0  1  0  0 -1  0
  0  1  0  0 -1  0
  0  1  0  0 -1  0
  0  0  0  0  0  0
[torch.FloatTensor of size 1x1x5x6]

This output is the result for cross-correlation. Therefore, we need to flip the filter

def flip_tensor(t):
    flipped = t.numpy().copy()

    for i in range(len(filters.size())):
        flipped = np.flip(flipped,i) #Reverse given tensor on dimention i
    return torch.from_numpy(flipped.copy())

print(F.conv2d(inputs, autograd.Variable(flip_tensor(filters.data))))

The new output is the correct result for convolution.

Variable containing:
(0 ,0 ,.,.) = 
  0  0  0  0  0  0
  0 -1  0  0  1  0
  0 -1  0  0  1  0
  0 -1  0  0  1  0
  0  0  0  0  0  0
[torch.FloatTensor of size 1x1x5x6]

来源：https://stackoverflow.com/questions/42970009/performing-convolution-not-cross-correlation-in-pytorch

标签

neural-network

conv-neural-network

pytorch

convolution

torch