- 🍨 本文为🔗365天深度学习训练营中的学习记录博客
- 🍖 原作者：K同学啊|接辅导、项目定制
一、课题背景和开发环境
📌第J4周：ResNet与DenseNet结合探索📌
- 语言：Python3、Pytorch
- 📌本周任务：📌
– 任务类型： 自主探索
– 任务难度： 偏难
– 任务描述：
Ⅰ.请根据J1~J3周的内容自由探索ResNet与DenseNet结合的可能性
Ⅱ.是否可以根据两种特性构建一个新的模型框架？
Ⅲ.请用之前的任一图像识别任务验证改进后的模型的效果
🔊注： 打卡内容应该包括创新的思路以及对应模型结构图、代码运行截图
二、网络结构
在网上找到了一个ResNet与DenseNet复合的网络框架DPN，本次课题就以实现DPN（Dual Path Networks）为主。
参考资料：
Higher Order Recurrent Neural Networks
Dual Path Networks
DPN详解（Dual Path Networks）
解读Dual Path Networks（DPN，原创）
Dual Path Networks双分支网络
pytorch实现DPN 最详细的全面讲解
三、使用Pytorch实现
pytorch
class Block(nn.Module):
def __init__(self, in_channel, mid_channel, out_channel, dense_channel, stride, groups, is_shortcut=False):
# in_channel,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方式的卷积运算，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
super(Block, self).__init__()
self.is_shortcut = is_shortcut
self.out_channel = out_channel
self.conv1 = nn.Sequential(
nn.Conv2d(in_channel, mid_channel, kernel_size=1, bias=False),
nn.BatchNorm2d(mid_channel),
nn.ReLU()
)
self.conv2 = nn.Sequential(
nn.Conv2d(mid_channel, mid_channel, kernel_size=3, stride=stride, padding=1, groups=groups, bias=False),
nn.BatchNorm2d(mid_channel),
nn.ReLU()
)
self.conv3 = nn.Sequential(
nn.Conv2d(mid_channel, out_channel+dense_channel, kernel_size=1, bias=False),
nn.BatchNorm2d(out_channel+dense_channel)
)
if self.is_shortcut:
self.shortcut = nn.Sequential(
nn.Conv2d(in_channel, out_channel+dense_channel, kernel_size=3, padding=1, stride=stride, bias=False),
nn.BatchNorm2d(out_channel+dense_channel)
)
self.relu = nn.ReLU(inplace=True)
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_channel
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__()
self.group = cfg['group']
self.in_channel = cfg['in_channel']
mid_channels = cfg['mid_channels']
out_channels = cfg['out_channels']
dense_channels = cfg['dense_channels']
num = cfg['num']
self.conv1 = nn.Sequential(
nn.Conv2d(3, self.in_channel, 7, stride=2, padding=3, bias=False, padding_mode='zeros'),
nn.BatchNorm2d(self.in_channel),
nn.ReLU(),
nn.MaxPool2d(kernel_size=3, stride=2, padding=0)
)
self.conv2 = self._make_layers(mid_channels[0], out_channels[0], dense_channels[0], num[0], stride=1)
self.conv3 = self._make_layers(mid_channels[1], out_channels[1], dense_channels[1], num[1], stride=2)
self.conv4 = self._make_layers(mid_channels[2], out_channels[2], dense_channels[2], num[2], stride=2)
self.conv5 = self._make_layers(mid_channels[3], out_channels[3], dense_channels[3], num[3], stride=2)
self.pool = nn.AdaptiveAvgPool2d((1, 1))
self.fc = nn.Linear(cfg['out_channels'][3] + (num[3]+1) * cfg['dense_channels'][3], cfg['classes']) # fc层需要计算
def _make_layers(self, mid_channel, out_channel, dense_channel, num, stride=2):
layers = []
layers.append(Block(self.in_channel, mid_channel, out_channel, dense_channel, stride=stride, groups=self.group, is_shortcut=True))
# 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
# 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
self.in_channel = out_channel + dense_channel*2
# 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_channel,后面每次一都会多出1倍，所以有(i+2)*dense_channel
for i in range(1, num):
layers.append(Block(self.in_channel, mid_channel, out_channel, dense_channel, stride=1, groups=self.group))
self.in_channel = self.in_channel + dense_channel
#self.in_channel = out_channel + (i+2)*dense_channel
return nn.Sequential(*layers)
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.pool(x)
x = torch.flatten(x, start_dim=1)
x = self.fc(x)
return x
def DPN92(n_class=10):
cfg = {
'group': 32,
'in_channel': 64,
'mid_channels': (96, 192, 384, 768),
'out_channels': (256, 512, 1024, 2048),
'dense_channels': (16, 32, 24, 128),
'num': (3, 4, 20, 3),
'classes': (n_class)
}
return DPN(cfg)
def DPN98(n_class=10):
cfg = {
'group': 40,
'in_channel': 96,
'mid_channels': (160, 320, 640, 1280),
'out_channels': (256, 512, 1024, 2048),
'dense_channels': (16, 32, 32, 128),
'num': (3, 6, 20, 3),
'classes': (n_class)
}
return DPN(cfg)
四、打印模型查看&运行
1.模型打印
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 112, 112] 9,408
BatchNorm2d-2 [-1, 64, 112, 112] 128
ReLU-3 [-1, 64, 112, 112] 0
MaxPool2d-4 [-1, 64, 55, 55] 0
Conv2d-5 [-1, 96, 55, 55] 6,144
BatchNorm2d-6 [-1, 96, 55, 55] 192
ReLU-7 [-1, 96, 55, 55] 0
Conv2d-8 [-1, 96, 55, 55] 2,592
BatchNorm2d-9 [-1, 96, 55, 55] 192
ReLU-10 [-1, 96, 55, 55] 0
Conv2d-11 [-1, 272, 55, 55] 26,112
BatchNorm2d-12 [-1, 272, 55, 55] 544
Conv2d-13 [-1, 272, 55, 55] 156,672
BatchNorm2d-14 [-1, 272, 55, 55] 544
ReLU-15 [-1, 288, 55, 55] 0
Block-16 [-1, 288, 55, 55] 0
Conv2d-17 [-1, 96, 55, 55] 27,648
BatchNorm2d-18 [-1, 96, 55, 55] 192
ReLU-19 [-1, 96, 55, 55] 0
Conv2d-20 [-1, 96, 55, 55] 2,592
BatchNorm2d-21 [-1, 96, 55, 55] 192
ReLU-22 [-1, 96, 55, 55] 0
Conv2d-23 [-1, 272, 55, 55] 26,112
BatchNorm2d-24 [-1, 272, 55, 55] 544
ReLU-25 [-1, 304, 55, 55] 0
Block-26 [-1, 304, 55, 55] 0
Conv2d-27 [-1, 96, 55, 55] 29,184
BatchNorm2d-28 [-1, 96, 55, 55] 192
ReLU-29 [-1, 96, 55, 55] 0
Conv2d-30 [-1, 96, 55, 55] 2,592
BatchNorm2d-31 [-1, 96, 55, 55] 192
ReLU-32 [-1, 96, 55, 55] 0
Conv2d-33 [-1, 272, 55, 55] 26,112
BatchNorm2d-34 [-1, 272, 55, 55] 544
ReLU-35 [-1, 320, 55, 55] 0
Block-36 [-1, 320, 55, 55] 0
Conv2d-37 [-1, 192, 55, 55] 61,440
BatchNorm2d-38 [-1, 192, 55, 55] 384
ReLU-39 [-1, 192, 55, 55] 0
Conv2d-40 [-1, 192, 28, 28] 10,368
BatchNorm2d-41 [-1, 192, 28, 28] 384
ReLU-42 [-1, 192, 28, 28] 0
Conv2d-43 [-1, 544, 28, 28] 104,448
BatchNorm2d-44 [-1, 544, 28, 28] 1,088
Conv2d-45 [-1, 544, 28, 28] 1,566,720
BatchNorm2d-46 [-1, 544, 28, 28] 1,088
ReLU-47 [-1, 576, 28, 28] 0
Block-48 [-1, 576, 28, 28] 0
Conv2d-49 [-1, 192, 28, 28] 110,592
BatchNorm2d-50 [-1, 192, 28, 28] 384
ReLU-51 [-1, 192, 28, 28] 0
Conv2d-52 [-1, 192, 28, 28] 10,368
BatchNorm2d-53 [-1, 192, 28, 28] 384
ReLU-54 [-1, 192, 28, 28] 0
Conv2d-55 [-1, 544, 28, 28] 104,448
BatchNorm2d-56 [-1, 544, 28, 28] 1,088
ReLU-57 [-1, 608, 28, 28] 0
Block-58 [-1, 608, 28, 28] 0
Conv2d-59 [-1, 192, 28, 28] 116,736
BatchNorm2d-60 [-1, 192, 28, 28] 384
ReLU-61 [-1, 192, 28, 28] 0
Conv2d-62 [-1, 192, 28, 28] 10,368
BatchNorm2d-63 [-1, 192, 28, 28] 384
ReLU-64 [-1, 192, 28, 28] 0
Conv2d-65 [-1, 544, 28, 28] 104,448
BatchNorm2d-66 [-1, 544, 28, 28] 1,088
ReLU-67 [-1, 640, 28, 28] 0
Block-68 [-1, 640, 28, 28] 0
Conv2d-69 [-1, 192, 28, 28] 122,880
BatchNorm2d-70 [-1, 192, 28, 28] 384
ReLU-71 [-1, 192, 28, 28] 0
Conv2d-72 [-1, 192, 28, 28] 10,368
BatchNorm2d-73 [-1, 192, 28, 28] 384
ReLU-74 [-1, 192, 28, 28] 0
Conv2d-75 [-1, 544, 28, 28] 104,448
BatchNorm2d-76 [-1, 544, 28, 28] 1,088
ReLU-77 [
最低0.47元/天 解锁文章
扫一扫
1952
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
