在pytorch中执行卷积(非互相关)

我有一个网络(https://github.com/TheAbhiKumar/tensorflow-value-iteration-networks),我试图在pytorch中实现(我对pytorch很新,但是,根本不是机器学习的新手) .

简而言之,我似乎无法弄清楚如何在pytorch中实现"纯"卷积.在张量流中,它可以像这样完成:

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

使用flipkernel函数:

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

如何在pytorch中完成类似的事情？

TLDR使用功能工具箱中的卷积torch.nn.fuctional.conv2d,而不是torch.nn.conv2d,并围绕垂直和水平轴翻转滤镜.

torch.nn.conv2d是网络的卷积层.因为学习了权重,所以如果使用互相关来实现权重并不重要,因为网络将只是学习内核的镜像版本(感谢@etarion进行此澄清).

torch.nn.fuctional.conv2d使用作为参数提供的输入和权重执行卷积,类似于示例中的tensorflow函数.我写了一个简单的测试,以确定是否像tensorflow函数一样,它实际上是执行互相关,并且有必要翻转滤波器以获得正确的卷积结果.

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))

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]

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))))

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]