pytorch实现DPN 最详细的全面讲解
DPN这个模型融入了三种基础模型，inception，resnet，densenet。有inception的宽度又有resnet的shortcut利用和densenet的浅层特征重复利用，可以看出它的强大。这个模型比较难掌握，需要你之前实践过前三种模型才可以理解本篇的代码。前三种模型我都有详细的讲解，inception详细讲解 resnet详细讲解 densenet详细讲解完全理解这三篇就可以轻松的掌握DPN模型。
torch.cat的使用
第一步要理解下面一行的代码
x = torch.cat([a[:,:d,:,:]+x[:,:d,:,:], a[:,d:,:,:], x[:,d:,:,:]], dim=1)
再讲这段代码的含义之前我要先讲解切片的计算，请看下面几个例子。
a = torch.rand((2,3,5,5))
print(a[:,:1,:,:])
#输出
tensor([[[[0.8158, 0.1786, 0.3600, 0.6222, 0.6837],
[0.9744, 0.9082, 0.5492, 0.6188, 0.2063],
[0.1767, 0.1156, 0.7501, 0.6518, 0.0286],
[0.9356, 0.6780, 0.8628, 0.2419, 0.3672],
[0.3171, 0.0869, 0.4242, 0.0131, 0.3955]]],
基本的切片操作很好理解，在卷积的理解是特征图输出一层。再看下面二个例子
import torch
a = torch.rand((2,3,5,5))
b = torch.rand((2,3,5,5))
d = torch.cat([a[:,:1,:,:]+b[:,:1,:,:]],dim=1)
print(d.size())
#输出
torch.Size([2, 1, 5, 5])
d = torch.cat([a[:,:1,:,:],b[:,:1,:,:]],dim=1)
print(d.size())
#输出
torch.Size([2, 2, 5, 5])
输出的size不同，第一个是数据相加，值大小的变化，第二个是拼接了一个特征图。通道数的改变。
import torch
a = torch.rand((2,3,5,5))
b = torch.rand((2,3,5,5))
e = torch.cat([a[:,:1,:,:]+b[:,:1,:,:],a[:,1:,:,:]],dim=1)
c= torch.cat([a[:,:1,:,:]+b[:,:1,:,:],a[:,1:,:,:], b[:,1:,:,:]],dim=1)
print(e.size())
print(c.size())
#输出
torch.Size([2, 3, 5, 5])
torch.Size([2, 5, 5, 5])
e的通道数为3，[a[:,:1,:,:]+b[:,:1,:,:],通道数不变，值得相加，后面a[:,1:,:,:]=a[:,1:3,:,:]这样看更明白，前面是一个通道数，后面是二个通道数，相加就是三个通道数。到这里应该很明白torch.cat的使用了吧。
融合的模型
DPN对特征图的利用非常的全面，一边有shortcut，又有dense浅层重复利用，所以构造起来也是相当难以理解。请看下面一个板块的构造。
class Block(nn.Module):
def __init__(self,in_channels,mid_channels,out_channels,dense_channels,stride,is_shortcut):
super(Block,self).__init__()
self.is_shortcut = is_shortcut
self.out_channels = out_channels
self.relu = nn.ReLU(inplace=True)
self.conv1 = nn.Sequential(
nn.Conv2d(in_channels,mid_channels,kernel_size=1,bias=False),
nn.BatchNorm2d(mid_channels),
nn.ReLU()
)
self.conv2 = nn.Sequential(
nn.Conv2d(mid_channels,mid_channels,kernel_size=3,stride=stride,padding=1,groups=32,bias=False),
nn.BatchNorm2d(mid_channels),
nn.ReLU()
)
self.conv3 = nn.Sequential(
nn.Conv2d(mid_channels,out_channels+dense_channels,kernel_size=1,bias=False),
nn.BatchNorm2d(out_channels+dense_channels)
)
if self.is_shortcut:
self.shortcut = nn.Sequential(
nn.Conv2d(in_channels,out_channels+dense_channels,kernel_size=1,stride=stride,bias=False),
nn.BatchNorm2d(out_channels+dense_channels)
)
def forward(self, x):
a = x
x = self.conv1(x)
x = self.conv2(x)
x = self.conv3(x)
if self.is_shortcut:
a = self.shortcut(a)
d = self.out_channels
x = torch.cat([a[:,:d,:,:]+x[:,:d,:,:], a[:,d:,:,:], x[:,d:,:,:]], dim=1)
x = self.relu(x)
return x
in_channels,是输入通道数,mid_channel是中间经历的通道数，out_channels是经过一次板块之后的输出通道数。dense_channels设置这个参数的原因就是一边进行着resnet方式的卷进运算，另一边也同时进行着dense的卷积计算，之后特征图融合形成新的特征图。
self.conv3 = nn.Sequential(
nn.Conv2d(mid_channels,out_channels+dense_channels,kernel_size=1,bias=False),
nn.BatchNorm2d(out_channels+dense_channels)
)
self.shortcut = nn.Sequential(
nn.Conv2d(in_channels,out_channels+dense_channels,kernel_size=1,stride=stride,bias=False),
nn.BatchNorm2d(out_channels+dense_channels)
核心点就是out_channels+dense_channels这个含义就是代表就同时进行resnet和dense卷积。
模型结构的讲解
def _make_layer(self,mid_channels,out_channels,dense_channels,num,stride):
layers = []
block_1 = Block(self.in_channels,mid_channels,out_channels,dense_channels,stride,is_shortcut=True)
self.in_channels = out_channels + 2*dense_channels
layers.append(block_1)
for i in range(1, num):
layers.append(Block(self.in_channels,mid_channels,out_channels,dense_channels,stride=1,is_shortcut=False))
self.in_channels = out_channels + (i+2)*dense_channels
return nn.Sequential(*layers)
block_1里面is_shortcut=True就是resnet中的shortcut连接，将浅层的特征进行一次卷积之后与进行三次卷积的特征图相加。 self.in_channels = out_channels + 2*dense_channels由于里面包含dense这种一直在叠加的特征图计算，所以第一次是2倍的dense_channels,每次一都会多出一倍，所以有(i+2)*dense_channels，后面几次相同的板块is_shortcut=False简单的理解就是一个多次重复的板块，第一次利用就可以满足浅层特征的利用。后面重复的不在需要。
对应的结构输出如下图
全部代码
import torch
import torch.nn as nn
class Block(nn.Module):
def __init__(self,in_channels,mid_channels,out_channels,dense_channels,stride,is_shortcut):
super(Block,self).__init__()
self.is_shortcut = is_shortcut
self.out_channels = out_channels
self.relu = nn.ReLU(inplace=True)
self.conv1 = nn.Sequential(
nn.Conv2d(in_channels,mid_channels,kernel_size=1,bias=False),
nn.BatchNorm2d(mid_channels),
nn.ReLU()
)
self.conv2 = nn.Sequential(
nn.Conv2d(mid_channels,mid_channels,kernel_size=3,stride=stride,padding=1,groups=32,bias=False),
nn.BatchNorm2d(mid_channels),
nn.ReLU()
)
self.conv3 = nn.Sequential(
nn.Conv2d(mid_channels,out_channels+dense_channels,kernel_size=1,bias=False),
nn.BatchNorm2d(out_channels+dense_channels)
)
if self.is_shortcut:
self.shortcut = nn.Sequential(
nn.Conv2d(in_channels,out_channels+dense_channels,kernel_size=1,stride=stride,bias=False),
nn.BatchNorm2d(out_channels+dense_channels)
)
def forward(self, x):
a = x
x = self.conv1(x)
x = self.conv2(x)
x = self.conv3(x)
if self.is_shortcut:
a = self.shortcut(a)
d = self.out_channels
x = torch.cat([a[:,:d,:,:]+x[:,:d,:,:], a[:,d:,:,:], x[:,d:,:,:]], dim=1)
x = self.relu(x)
return x
class DPN(nn.Module):
def __init__(self,cfg):
super(DPN,self).__init__()
mid_channels = cfg['mid_channels']
out_channels = cfg['out_channels']
num = cfg['num']
dense_channels = cfg['dense_channels']
self.in_channels = 64
self.conv1 = nn.Sequential(
nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False),
nn.BatchNorm2d(64),
nn.ReLU(),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
)
self.conv2 = self._make_layer(mid_channels[0], out_channels[0], dense_channels[0], num[0], stride=1)
self.conv3 = self._make_layer(mid_channels[1], out_channels[1], dense_channels[1], num[1], stride=2)
self.conv4 = self._make_layer(mid_channels[2], out_channels[2], dense_channels[2], num[2], stride=2)
self.conv5 = self._make_layer(mid_channels[3], out_channels[3], dense_channels[3], num[3], stride=2)
self.global_average_pool = nn.AdaptiveAvgPool2d((1,1))
self.fc = nn.Linear(out_channels[3]+(num[3]+1)*dense_channels[3], cfg['classes'])
def forward(self, x):
x = self.conv1(x)
x = self.conv2(x)
x = self.conv3(x)
x = self.conv4(x)
x = self.conv5(x)
x = self.global_average_pool(x)
x = torch.flatten(x, 1)
x = self.fc(x)
return x
def _make_layer(self,mid_channels,out_channels,dense_channels,num,stride):
layers = []
block_1 = Block(self.in_channels,mid_channels,out_channels,dense_channels,stride,is_shortcut=True)
self.in_channels = out_channels + 2*dense_channels
layers.append(block_1)
for i in range(1, num):
layers.append(Block(self.in_channels,mid_channels,out_channels,dense_channels,stride=1,is_shortcut=False))
self.in_channels = out_channels + (i+2)*dense_channels
return nn.Sequential(*layers)
def DPN92():
cfg = {
'mid_channels': (96,192,384,768),
'out_channels': (256,512,1024,2048),
'num': (3,4,20,3),
'dense_channels': (16,32,24,128),
'classes': (10)
}
return DPN(cfg)
net = DPN92()
x = torch.rand((10, 3, 224, 224))
for name,layer in net.named_children():
if name != "fc":
x = layer(x)
print(name, 'output shaoe:', x.shape)
else:
x = x.view(x.size(0), -1)
x = layer(x)
print(name, 'output shaoe:', x.shape)