Pytorch卷积groups参数

可以看出通过groups划分卷积核参数量的变化：

from torch import nn

#这个在 depthwise 中的效果
t = nn.Conv2d(10, 10, 3, 2, 0, bias=False, groups=10)
t.weight.shape
#torch.Size([10, 1, 3, 3])

t = nn.Conv2d(10, 10, 3, 2, 0, bias=False, groups=2)
t.weight.shape
#torch.Size([10, 5, 3, 3])

t = nn.Conv2d(10, 10, 3, 2, 0, bias=False, groups=1)
t.weight.shape
#torch.Size([10, 10, 3, 3])

下面是卷积结果和我们手工计算结果：结果是一样的

import numpy as np
import torch
from torch import nn

input_feature_size = 5

t = nn.Conv2d(input_feature_size, input_feature_size, 3, 2, 0, bias=False, groups=input_feature_size)
print(t.weight.shape)
w0 = t.weight[0, 0]
feature_map = torch.ones([input_feature_size, 3, 3])

ret = t(feature_map)

print("torch 卷积结果：", ret)
my_ret = []
for one_w in t.weight:
    sum_ = 0
    for one_fm in feature_map:
        sum_ += torch.sum(one_fm * one_w[0])
    my_ret.append(sum_/input_feature_size)
print("手工计算 卷积结果：", ret)
print('end!')

下面这个是普通卷积操作，groups=1：

import numpy as np
import torch
from torch import nn

input_feature_size = 5

t = nn.Conv2d(input_feature_size, input_feature_size, 3, 2, 0, bias=False, groups=1)
print(t.weight.shape)
w0 = t.weight[0, 0]
feature_map = torch.ones([input_feature_size, 3, 3])

ret = t(feature_map)

print("torch 卷积结果：", ret)
my_ret = []
for one_w in t.weight:
    sum_ = 0
    for one_fm in feature_map:
        sum_ += torch.sum(one_fm * one_w)
    my_ret.append(sum_)
print("手工计算 卷积结果：", ret)
print('end!')

可以看出在groups非1是卷积结果是需要除以featuremap的个数的。