pytorch之分组卷积
分组卷积
推荐：请先看最底部的参考连接
普通卷积
说明：普通卷积为组数为1的特殊分组卷积。
class Conv(nn.Module):
def __init__(self, in_ch, out_ch, groups):
super(Conv, self).__init__()
self.conv = nn.Conv2d(
in_channels=in_ch,
out_channels=out_ch,
kernel_size=3,
stride=1,
padding=1,
groups=groups
)
def forward(self, input):
out = self.conv(input)
return out
if __name__ == "__main__":
net = Group(64, 256, 1)
summary(net, (64, 224, 224), 1, "cpu")
---------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [1, 256, 224, 224] 147,712
================================================================
Total params: 147,712
Trainable params: 147,712
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 12.25
Forward/backward pass size (MB): 98.00
Params size (MB): 0.56
Estimated Total Size (MB): 110.81
----------------------------------------------------------------
分组卷积
说明：可以人为指定分组大小，理论上减少的参数位1/G（G为分组数目）。
class Group(nn.Module):
def __init__(self, in_ch, out_ch, groups):
super(Group, self).__init__()
self.conv = nn.Conv2d(
in_channels=in_ch,
out_channels=out_ch,
kernel_size=3,
stride=1,
padding=1,
groups=groups
)
def forward(self, input):
out = self.conv(input)
return out
if __name__ == "__main__":
net = Group(64, 256, 8)
summary(net, (64, 224, 224), 1, "cpu")
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [1, 256, 224, 224] 18,688
================================================================
Total params: 18,688
Trainable params: 18,688
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 12.25
Forward/backward pass size (MB): 98.00
Params size (MB): 0.07
Estimated Total Size (MB): 110.32
----------------------------------------------------------------
减少的参数
# 理论上减少的参数位１/G
print("理论值：{}\n实际值：{}".format(1/8, 18688/147712))
理论值：0.125
实际值：0.1265164644714038
深度可分离卷积
说明：组数等于输入通道数。。
class DepthwiseSeparableConv(nn.Module):
def __init__(self, in_ch, out_ch):
super(DepthwiseSeparableConv, self).__init__()
# 分组数目等于通道数目
self.depth_conv = nn.Conv2d(
in_channels=in_ch,
out_channels=in_ch,
kernel_size=3,
stride=1,
padding=1,
groups=in_ch
)
# 使用1*1卷积得到指定的输出通道
self.point_conv = nn.Conv2d(
in_channels=in_ch,
out_channels=out_ch,
kernel_size=1,
stride=1,
padding=0,
groups=1
)
def forward(self, input):
out = self.depth_conv(input)
out = self.point_conv(out)
return out
if __name__ == "__main__":
net = DepthwiseSeparableConv(64, 256)
summary(net, (64, 224, 224), 1, "cpu")
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [1, 64, 224, 224] 640
Conv2d-2 [1, 256, 224, 224] 16,640
================================================================
Total params: 17,280
Trainable params: 17,280
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 12.25
Forward/backward pass size (MB): 122.50
Params size (MB): 0.07
Estimated Total Size (MB): 134.82
----------------------------------------------------------------
应用网络（MobileFaceNet）
class DepthwiseSeparableConv(nn.Module):
def __init__(self, in_planes, out_planes, kernel_size, padding, bias=False):
super(DepthwiseSeparableConv, self).__init__()
self.depthwise = nn.Conv2d(in_planes, in_planes
, kernel_size=kernel_siz
e, padding=padding
, groups=in_planes
, bias=bias
)
self.pointwise = nn.Conv2d(in_planes, out_planes
, kernel_size=1
, bias=bias)
self.bn1 = nn.BatchNorm2d(in_planes)
self.bn2 = nn.BatchNorm2d(out_planes)
self.relu = nn.ReLU()
def forward(self, x):
x = self.depthwise(x)
x = self.bn1(x)
x = self.relu(x)
x = self.pointwise(x)
x = self.bn2(x)
x = self.relu(x)
return x
if __name__ == "__main__":
net = DepthwiseSeparableConv(64, 256, 3, 0)
summary(net, (64, 224, 224), 1, "cpu")
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [1, 64, 222, 222] 576
BatchNorm2d-2 [1, 64, 222, 222] 128
ReLU-3 [1, 64, 222, 222] 0
Conv2d-4 [1, 256, 222, 222] 16,384
BatchNorm2d-5 [1, 256, 222, 222] 512
ReLU-6 [1, 256, 222, 222] 0
================================================================
Total params: 17,600
Trainable params: 17,600
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 12.25
Forward/backward pass size (MB): 360.97
Params size (MB): 0.07
Estimated Total Size (MB): 373.28
----------------------------------------------------------------
全局可分离卷积
说明：组数等于输入通道，且卷积核大小等于输入大小。
class GDConv(nn.Module):
def __init__(self, in_planes, out_planes, kernel_size, padding, bias=False):
super(GDConv, self).__init__()
self.depthwise = nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, padding=padding, groups=in_planes,
bias=bias)
self.bn = nn.BatchNorm2d(in_planes)
def forward(self, x):
x = self.depthwise(x)
x = self.bn(x)
return x
if __name__ == "__main__":
net = GDConv(64, 64, 224, 0)
summary(net, (64, 224, 224), 1, "cpu")
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [1, 64, 1, 1] 3,211,264
BatchNorm2d-2 [1, 64, 1, 1] 128
================================================================
Total params: 3,211,392
Trainable params: 3,211,392
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 12.25
Forward/backward pass size (MB): 0.00
Params size (MB): 12.25
Estimated Total Size (MB): 24.50
----------------------------------------------------------------
Pytorch参数说明
groups – Number of blocked connections from input channels to output
- At groups=1, all inputs are convolved to all outputs.
- At groups=2, the operation becomes equivalent to having two conv layers side by side, each seeing half the input channels, and producing half the output channels, and both subsequently concatenated.
- At groups= in_channels, each input channel is convolved with its own set of filters, of size: ⌊(out_channels)/(in_channels)⌋.
参考
分组卷积图片来源
Group Convolution分组卷积，以及Depthwise Convolution和Global Depthwise Convolution
深度可分离卷积(Depthwise Separable Convolution)和分组卷积(Group Convolution)的理解，相互关系及PyTorch实现
MobileFaceNet