（原文）网络结构：
（原文）详细结构
（原文）插图：
"""
2021年09月27日21:56:08:DenseNet完成, 细节其实可以修改下,
尤其是一个block中, 维度的合并,以及1*1降维的维数可以指定一下
简易版本：自己按照论文以自己的理解写的，有不足之处还请见谅，可在评论区交流指正
"""
import os
import torch
import torch.nn as nn
os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
class DenseNet(nn.Module):
def __init__(self, n_class):
super(DenseNet, self).__init__()
self.layer1 = nn.Sequential(
nn.Conv2d(3, 3, 7, stride=2, padding=4),
nn.MaxPool2d(3, 2),
)
self.layer_dense2 = DenseBlock(3)
self.transition2 = self.Transition_Layer(15, 6)
self.layer_dense3 = DenseBlock(6)
self.transition3 = self.Transition_Layer(30, 12)
self.layer_dense4 = DenseBlock(12)
self.transition4 = self.Transition_Layer(60, 120)
self.layer_pool5 = nn.AvgPool2d(7, 7)
self.Linear = nn.Linear(120, n_class)
def Transition_Layer(self, in_, out):
"""
控制升维和降维
:param in_:
:param out:
:return:
"""
transition = nn.Sequential(
nn.BatchNorm2d(in_),
nn.ReLU(),
nn.Conv2d(in_, out, 1),
nn.AvgPool2d(2, 2),
)
return transition
def forward(self, x):
x = self.layer1(x)
x = self.layer_dense2(x)
x = self.transition2(x)
x = self.layer_dense3(x)
x = self.transition3(x)
x = self.layer_dense4(x)
x = self.transition4(x)
x = self.layer_pool5(x)
x = x.view(x.size(0), -1)
x = self.Linear(x)
return x
class DenseBlock(nn.Module):
def __init__(self, in_channel):
"""
每次卷积输入输出的模块维度是相同的, 最后拼接在一起
:param in_channel:输入维度, 输出维度相同
"""
super(DenseBlock, self).__init__()
self.d1 = self.Conv_Block(in_channel, in_channel)
self.d2 = self.Conv_Block(2 * in_channel, in_channel)
self.d3 = self.Conv_Block(4 * in_channel, in_channel)
self.d4 = self.Conv_Block(8 * in_channel, in_channel)
@staticmethod
def Conv_Block(in_channel, out):
Conv = nn.Sequential(
nn.BatchNorm2d(in_channel),
nn.ReLU(),
nn.Conv2d(in_channel, out, 1),
nn.BatchNorm2d(out),
nn.ReLU(),
nn.Conv2d(out, out, 3, padding=1),
)
return Conv
def forward(self, x):
x1 = self.d1(x)
x_cat1 = torch.cat((x, x1), dim=1)
x2 = self.d2(x_cat1)
x_cat2 = torch.cat((x2, x_cat1, x1), dim=1)
x3 = self.d3(x_cat2)
x_cat3 = torch.cat((x3, x_cat2, x_cat1, x1), dim=1)
x4 = self.d4(x_cat3)
x = torch.cat((x4, x3, x2, x1, x), dim=1)
return x
if __name__ == '__main__':
inputs = torch.randn(10, 3, 224, 224)
model = DenseNet(n_class=6)
outputs = model(inputs)
print(model)
print(outputs.shape)
