Pytorch Conv2d 中的group测试

测试Pytorch Conv2d 中的group参数实际影响：
首先定义一个我们能验算的3 33的数组

import torch
import numpy as np
array1 = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9]).reshape(3, 3)
array2 = np.array([2, 3, 4, 5, 6, 7, 8, 9, 1]).reshape(3, 3)
array3 = np.array([3, 4, 5, 6, 7, 8, 9, 1, 2]).reshape(3, 3)

array1 = array1[np.newaxis, :, :]
array2 = array1[np.newaxis, :, :]
array3 = array1[np.newaxis, :, :]
array = np.vstack((array1, array2, array3))
array = array[np.newaxis, :, :, :]
array = torch.tensor(array).float()

print(array)

tensor([[[[1., 2., 3.],
          [4., 5., 6.],
          [7., 8., 9.]],

         [[2., 3., 4.],
          [5., 6., 7.],
          [8., 9., 1.]],

         [[3., 4., 5.],
          [6., 7., 8.],
          [9., 1., 2.]]]])

然后定义两个group不同的卷积层，一个group=1，另一个group=3

conv1 = torch.nn.Conv2d(3, 6, 1, 1, autopad(1), groups=1, bias=False)

conv2 = torch.nn.Conv2d(3, 6, 1, 1, autopad(1), groups=3, bias=False)

第一个卷积输出

print('conv1.weight', conv1.weight)
print('(conv1(array)', conv1(array))

conv1.weight Parameter containing:
tensor([[[[-0.21770]],

         [[ 0.55208]],

         [[-0.19912]]],


        [[[-0.10678]],

         [[-0.11752]],

         [[ 0.56264]]],


        [[[ 0.48454]],

         [[-0.29961]],

         [[-0.00285]]],


        [[[-0.43829]],

         [[-0.42696]],

         [[-0.34118]]],


        [[[ 0.03800]],

         [[ 0.36227]],

         [[ 0.50197]]],


        [[[-0.40956]],

         [[-0.28049]],

         [[-0.14078]]]], requires_grad=True)
         
         (conv1(array) tensor([[[[ 0.28909,  0.42435,  0.55960],
          [ 0.69486,  0.83012,  0.96537],
          [ 1.10063,  3.02800, -1.80544]],

         [[ 1.34611,  1.68446,  2.02280],
          [ 2.36115,  2.69949,  3.03784],
          [ 3.37618, -1.34927,  0.04675]],

         [[-0.12323,  0.05885,  0.24093],
          [ 0.42301,  0.60509,  0.78717],
          [ 0.96925,  1.17697,  4.05551]],

         [[-2.31577, -3.52222, -4.72866],
          [-5.93510, -7.14154, -8.34798],
          [-9.55442, -7.69021, -5.05398]],

         [[ 2.26845,  3.17069,  4.07293],
          [ 4.97516,  5.87740,  6.77964],
          [ 7.68188,  4.06638,  1.70823]],

         [[-1.39288, -2.22370, -3.05453],
          [-3.88535, -4.71618, -5.54700],
          [-6.37783, -5.94164, -4.24805]]]], grad_fn=<ThnnConv2DBackward>)

验算第一个元素
0.28909 = -0.21770 + 2 * 0.55208 - 3 * 0.19912
可知这个卷积是核对于三层相乘之和。

第二个卷积输出

print('conv2.weight', conv2.weight)
print('(conv2(array)', (conv2(array)))

conv2.weight Parameter containing:
tensor([[[[ 0.06971]]],


        [[[ 0.50235]]],


        [[[-0.92221]]],


        [[[-0.07961]]],


        [[[-0.02840]]],


        [[[ 0.12154]]]], requires_grad=True)
(conv2(array) tensor([[[[ 0.06971,  0.13942,  0.20912],
          [ 0.27883,  0.34854,  0.41825],
          [ 0.48795,  0.55766,  0.62737]],

         [[ 0.50235,  1.00471,  1.50706],
          [ 2.00941,  2.51176,  3.01412],
          [ 3.51647,  4.01882,  4.52117]],

         [[-1.84443, -2.76664, -3.68885],
          [-4.61107, -5.53328, -6.45549],
          [-7.37771, -8.29992, -0.92221]],

         [[-0.15921, -0.23882, -0.31843],
          [-0.39803, -0.47764, -0.55725],
          [-0.63686, -0.71646, -0.07961]],

         [[-0.08519, -0.11359, -0.14199],
          [-0.17039, -0.19879, -0.22718],
          [-0.25558, -0.02840, -0.05680]],

         [[ 0.36461,  0.48615,  0.60768],
          [ 0.72922,  0.85076,  0.97229],
          [ 1.09383,  0.12154,  0.24307]]]], grad_fn=<MkldnnConvolutionBackward>)

肉眼观察可知，
结果的第1，2层是
矩阵

[[1., 2., 3.],
[4., 5., 6.],
[7., 8., 9.]],

乘以核0.06971和 0.50235的结果，第3，4，5，6层以此类推。
可知，输入的3层被分了3组，每组与一个核相乘。

总结

我个人的理解是，对于输入是图片的信息来看，一张图片的三个通道的信息会因为卷积完全融到一块，打破了层与层之间的信息壁垒，而分组之后的卷积相当于只对某一层进行处理。