论文链接: CSPNet: A New Backbone that can Enhance Learning Capability of CNN
论文提出Cross Stage Partial(CSP)结构,其初衷是减少计算量并且增强梯度的表现。主要思想是:在输入block之前,将输入分为两个部分,其中一部分通过block进行计算,另一部分直接通过一个shortcut进行concatenate。
作者基于Dense block(对DenseNet结构不了解的同学可以查看我以前的博客:【深度学习基础】pytorch实现DenseNet亲身实践)讲述了CSP的结构,如下图:
图片来自于论文
上图中,上者为原始的Dense Block,下者为Partial Dense Block。假设将输入按照part_ratio=0.5的比例分成两部分,并且假设一个Dense Block的输入为whc,growth rate为k,block layers为m,则:对于原始的DenseBlock,其CIO为(Convolutional input/Output)为(cm) + ((mm+m)d)/2;而对于使用了CSP结构的结构来说,其CIO下降为(cm) + ((m*m+m)*d)/2。
作者在论文阐述了CSP结构的优点:
(1)加强CNN的学习能力;(2)减少计算瓶颈,现在的网络大多计算代价昂贵,不利于工业的落地;(3)减少内存消耗。
关于Patial Transition Layer,论文提出了三种不同的结构,如下图:
图片来自论文
Transition layer的含义同DenseNet,是一组1x1的卷积层。上图中,不同类型的transition layer决定了梯度的结构方式,并且各有其优势:(c)图Fusion First的先将两个部分进行concatenate,然后再进行输入到Transion layer中,采用这种做法会是的大量特梯度信息被重用,有利于网络学习;(d)图中Fusion Last先将部分特征输入Transition layer,然后再进行concatenate,这样做损失了部分的梯度重用,但是由于Transition的输入维度比(c)图少,大大减少计算复杂度。论文中CSPNet采用的是图(b)中的结构,其结合了(c)(d)的结合,进一步提升了学习能力,但是也进一步提高了一些计算复杂度。 作者在论文中给出其使用不同Partial Transition Layer的实验结果,如下图。具体使用哪种结构可以根据条件和使用场景进行调整。
图片来自论文
CSPNet沿用网络的Architecture,只是对于基础block进行改造,根据之前博客介绍的DenseNet结构修改相应代码即可实现。
沿用之前介绍DenseNet写的Dense Block代码:
class DenseBlock(nn.Module):
def __init__(self, input_channels, num_layers, growth_rate):
super(DenseBlock, self).__init__()
self.num_layers = num_layers
self.k0 = input_channels
self.k = growth_rate
self.layers = self.__make_layers()
def __make_layers(self):
layer_list = []
for i in range(self.num_layers):
layer_list.append(nn.Sequential(
BN_Conv2d(self.k0 + i * self.k, 4 * self.k, 1, 1, 0),
BN_Conv2d(4 * self.k, self.k, 3, 1, 1)
))
return layer_list
def forward(self, x):
feature = self.layers[0](x)
out = torch.cat((x, feature), 1)
for i in range(1, len(self.layers)):
feature = self.layers[i](out)
out = torch.cat((feature, out), 1)
return out
Partial Dense Block的实现:
class CSP_DenseBlock(nn.Module):
def __init__(self, in_channels, num_layers, k, part_ratio=0.5):
super(CSP_DenseBlock, self).__init__()
self.part1_chnls = int(in_channels * part_ratio)
self.part2_chnls = in_channels - self.part1_chnls
self.dense = DenseBlock(self.part2_chnls, num_layers, k)
# trans_chnls = self.part2_chnls + k * num_layers
# self.transtion = BN_Conv2d(trans_chnls, trans_chnls, 1, 1, 0)
def forward(self, x):
part1 = x[:, :self.part1_chnls, :, :]
part2 = x[:, self.part1_chnls:, :, :]
part2 = self.dense(part2)
# part2 = self.transtion(part2)
out = torch.cat((part1, part2), 1)
return out
上面的代码采用的是Fusion Last的Partial transition layer,取消其中的三行注释就是论文中采用的Partial transition layer。由于,我之前的Transition layer是写在整体网络结构中,没有放在block中,所以就不实现Fusion first的形式了。
在之前原始的DenseNet结构的基础上进行修改,增加一个partion-ratio参数来控制使用的block结构和通道划分的比例就可以了,代码:
class DenseNet(nn.Module):
def __init__(self, layers: object, k, theta, num_classes, part_ratio=0) -> object:
super(DenseNet, self).__init__()
# params
self.layers = layers
self.k = k
self.theta = theta
self.Block = DenseBlock if part_ratio == 0 else CSP_DenseBlock # 通过part_tatio参数控制block type
# layers
self.conv = BN_Conv2d(3, 2 * k, 7, 2, 3)
self.blocks, patches = self.__make_blocks(2 * k)
self.fc = nn.Linear(patches, num_classes)
def __make_transition(self, in_chls):
out_chls = int(self.theta * in_chls)
return nn.Sequential(
BN_Conv2d(in_chls, out_chls, 1, 1, 0),
nn.AvgPool2d(2)
), out_chls
def __make_blocks(self, k0):
"""
make block-transition structures
:param k0:
:return:
"""
layers_list = []
patches = 0
for i in range(len(self.layers)):
layers_list.append(self.Block(k0, self.layers[i], self.k))
patches = k0 + self.layers[i] * self.k # output feature patches from Dense Block
if i != len(self.layers) - 1:
transition, k0 = self.__make_transition(patches)
layers_list.append(transition)
return nn.Sequential(*layers_list), patches
def forward(self, x):
out = self.conv(x)
out = F.max_pool2d(out, 3, 2, 1)
# print(out.shape)
out = self.blocks(out)
# print(out.shape)
out = F.avg_pool2d(out, 7)
# print(out.shape)
out = out.view(out.size(0), -1)
out = F.softmax(self.fc(out))
return out
def csp_densenet_121(num_classes=1000):
return DenseNet([6, 12, 24, 16], k=32, theta=0.5, num_classes=num_classes, part_ratio=0.5)
def csp_densenet_169(num_classes=1000):
return DenseNet([6, 12, 32, 32], k=32, theta=0.5, num_classes=num_classes, part_ratio=0.5)
def csp_densenet_201(num_classes=1000):
return DenseNet([6, 12, 48, 32], k=32, theta=0.5, num_classes=num_classes, part_ratio=0.5)
def csp_densenet_264(num_classes=1000):
return DenseNet([6, 12, 64, 48], k=32, theta=0.5, num_classes=num_classes, part_ratio=0.5)
论文也给出了CSP-ResNet结构。我们知道在ResNeXt的实现中,若groups设置为1,就和ResNet结构相似,这里仅搭建CSP-ResNeXt。
图片来自论文
讨论:论文中提到由于输入的通道减半了,CSPNet不需要使用bottleneck结构了。但是我查看DarkNet框架下的网络参数,发现还是1x1,3x3,1x1的组合结构。有点纳闷,不知道我是不是对bottleneck结构有误解。可能作者指的是bottleneck指的是输出通道比输入通道数量要少的卷积层。
作者给的code链接是darknet框架下的cfg参数,由于我不熟悉这个框架,有点看不清楚。这里,我是按照论文的思想和自己理解搭建CSP-ResNeXt,网络输入尺寸和整体结构仍沿用ResNeXt的,不是论文实验确实采用的确切网络结构,谨供参考和了解思路。
我们再回顾一下ResNeXt的整体结构,之前博客的传送门:
图片来自ResNeXt论文
ResNeXt的主题是一些不同尺寸的block组成的,共4组,以ResNeXt-53为例,每组blocks的数量分别为[3, 4, 6, 3]。根据CSPNet的思想去除bottleneck结构,采用Fusion First和Fusion Last的模式。针对每一组,可以设计如下的结构:
图片来源:自己画的~
由上图,之前ResNeXt中Residual block中的最后一层1x1滤波器数量进行了2倍的expansion,这里改成三层均是一样的数量,那么我们根据这个结构进行编码。
再次提醒,这个复现不是SoAT算法所用的确切CSPResNet。但是其block结构符合上图中CSPNet的基本思想,整体架构沿用ResNet架构——虽然目前不准备实验验证,但我想它应该具有一定精度。
参考作者,激活函数采用了LeakyReLU,但在一些目标检测框架中会用Mish作为激活函数,以更高的计算代价获取网络更好的收敛性。
class BN_Conv2d_Leaky(nn.Module):
"""
BN_CONV_LeakyRELU
"""
def __init__(self, in_channels: object, out_channels: object, kernel_size: object, stride: object, padding: object,
dilation=1, groups=1, bias=False) -> object:
super(BN_Conv2d_Leaky, self).__init__()
self.seq = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride,
padding=padding, dilation=dilation, groups=groups, bias=bias),
nn.BatchNorm2d(out_channels)
)
def forward(self, x):
return F.leaky_relu(self.seq(x))
在这里,将最后一层1x1的数量改为和之前两层一致。在shortcut的实现上,我用均值池化来应对stride不为1的情况,因为感觉我之前在ResNeXt写的在1x1卷积的时候设置stride处理方式会损失一定的输入信息。 代码如下:
class ResidualBlock(nn.Module):
"""
Residual block for CSP-ResNeXt
"""
def __init__(self, in_channels, cardinality, group_width, stride=1):
super(ResidualBlock, self).__init__()
self.out_channels = cardinality * group_width
self.conv1 = BN_Conv2d_Leaky(in_channels, self.out_channels, 1, 1, 0)
self.conv2 = BN_Conv2d_Leaky(self.out_channels, self.out_channels, 3, stride, 1, groups=cardinality)
self.conv3 = nn.Conv2d(self.out_channels, self.out_channels, 1, 1, 0)
self.bn = nn.BatchNorm2d(self.out_channels)
# make shortcut
layers = []
if in_channels != self.out_channels:
layers.append(nn.Conv2d(in_channels, self.out_channels, 1, 1, 0))
layers.append(nn.BatchNorm2d(self.out_channels))
if stride != 1:
layers.append(nn.AvgPool2d(stride))
self.shortcut = nn.Sequential(*layers)
def forward(self, x):
out = self.conv3(self.conv2(self.conv1(x)))
out = self.bn(out)
out += self.shortcut(x)
return F.leaky_relu(out)
1、把上图中设计的结构作为一个整体来实现,称之为Stem,代码如下:
class Stem(nn.Module):
def __init__(self, in_channels, num_blocks, cardinality, group_with, stride=2):
super(Stem, self).__init__()
self.c0 = in_channels // 2
self.c1 = in_channels - in_channels // 2
self.hidden_channels = cardinality * group_with
self.out_channels = self.hidden_channels * 2
self.trans_part0 = nn.Sequential(BN_Conv2d_Leaky(self.c0, self.hidden_channels, 1, 1, 0), nn.AvgPool2d(stride))
self.block = self.__make_block(num_blocks, self.c1, cardinality, group_with, stride)
self.trans_part1 = BN_Conv2d_Leaky(self.hidden_channels, self.hidden_channels, 1, 1, 0)
self.trans = BN_Conv2d_Leaky(self.out_channels, self.out_channels, 1, 1, 0)
def __make_block(self, num_blocks, in_channels, cardinality, group_with, stride):
strides = [stride] + [1] * (num_blocks-1)
channels = [in_channels] + [self.hidden_channels] * (num_blocks-1)
return nn.Sequential(*[ResidualBlock(c, cardinality, group_with, s)
for c, s in zip(channels, strides)])
def forward(self, x):
x0 = x[:, :self.c0, :, :]
x1 = x[:, self.c0:, :, :]
out0 = self.trans_part0(x0)
out1 = self.trans_part1(self.block(x1))
out = torch.cat((out0, out1), 1)
return self.trans(out)
2、以下代码提供了CSP-ResNeXt的架构以及CSP-ResNeXt-53(32x4d)的接口:
class CSP_ResNeXt(nn.Module):
def __init__(self, num_blocks, cadinality, group_width, num_classes):
super(CSP_ResNeXt, self).__init__()
self.conv0 = BN_Conv2d_Leaky(3, 64, 7, 2, 3)
self.pool1 = nn.MaxPool2d(3, 2, 1)
self.conv1 = BN_Conv2d_Leaky(64, 128, 1, 1, 0)
self.stem0 = Stem(cadinality*group_width*2, num_blocks[0], cadinality, group_width, stride=1)
self.stem1 = Stem(cadinality*group_width*4, num_blocks[1], cadinality, group_width*2)
self.stem2 = Stem(cadinality*group_width*8, num_blocks[2], cadinality, group_width*4)
self.stem3 = Stem(cadinality*group_width*16, num_blocks[3], cadinality, group_width*8)
self.global_pool = nn.AdaptiveAvgPool2d((1, 1))
self.fc = nn.Linear(cadinality*group_width*16, num_classes)
def forward(self, x):
out = self.conv0(x)
out = self.pool1(out)
out = self.conv1(out)
out = self.stem0(out)
out = self.stem1(out)
out = self.stem2(out)
out = self.stem3(out)
out = self.global_pool(out)
out = out.view(out.size(0), -1)
out = self.fc(out)
return F.softmax(out)
def csp_resnext_50_32x4d(num_classes=1000):
return CSP_ResNeXt([3, 4, 6, 3], 32, 4, num_classes)
3、搭建完毕,测试网络:
net = csp_resnext_50_32x4d()
summary(net, (3, 256, 256))