- 🍨 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
本周任务：
- 探索ResNet和DenseNet的结合可能性
- 本周任务较难，我们在chatGPT的帮助下完成
一、网络的构建
设计一种结合 ResNet 和 DenseNet 的网络架构，目标是在性能与复杂度之间实现平衡，同时保持与 DenseNet-121 相当的训练速度，可以通过以下步骤设计一种新的网络结构，称为 ResDenseNet（暂命名）。这种网络结构结合了 ResNet 的残差连接和 DenseNet 的密集连接优点，同时对复杂度加以控制。
设计思路
残差模块与密集模块结合：
在网络的不同阶段，使用残差模块（ResBlock）来捕获浅层特征。
在每个阶段的后期引入密集模块（DenseBlock），实现高效的特征复用。
通过调整每层的通道数，避免过多的计算和内存消耗。
瓶颈设计（Bottleneck Block）：
每个模块采用瓶颈层，减少计算复杂度。
通过 1x1 卷积压缩和扩展特征通道数。
混合连接方式：
引入 局部密集连接，只连接同一模块内的层，避免 DenseNet 的全连接导致的内存开销。
在模块之间使用残差连接，便于信息流通。
网络深度与宽度的平衡：
将 DenseNet 的增长率（growth rate）减少，适当减少特征图通道数增长。
模块之间引入过渡层（Transition Layer）以压缩特征图尺寸和通道数。
import torch
import torch.nn as nnclass Bottleneck(nn.Module):def __init__(self, in_channels, growth_rate):super(Bottleneck, self).__init__()self.bn1 = nn.BatchNorm2d(in_channels)self.conv1 = nn.Conv2d(in_channels, 4 * growth_rate, kernel_size=1, stride=1, bias=False)self.bn2 = nn.BatchNorm2d(4 * growth_rate)self.conv2 = nn.Conv2d(4 * growth_rate, growth_rate, kernel_size=3, stride=1, padding=1, bias=False)def forward(self, x):out = self.conv1(self.bn1(x))out = self.conv2(self.bn2(out))return torch.cat([x, out], dim=1)class DenseBlock(nn.Module):def __init__(self, num_layers, in_channels, growth_rate):super(DenseBlock, self).__init__()self.layers = nn.ModuleList()for i in range(num_layers):self.layers.append(Bottleneck(in_channels + i * growth_rate, growth_rate))# 为了残差连接，可能需要调整通道数以匹配输入输出self.residual = nn.Conv2d(in_channels, in_channels + num_layers * growth_rate, kernel_size=1, bias=False)def forward(self, x):identity = self.residual(x) # 将输入调整为与 DenseBlock 输出通道一致for layer in self.layers:x = layer(x) # 密集连接，逐层拼接return x + identity # 残差连接：输入与输出相加class TransitionLayer(nn.Module):def __init__(self, in_channels, out_channels):super(TransitionLayer, self).__init__()self.bn = nn.BatchNorm2d(in_channels)self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, bias=False)self.pool = nn.AvgPool2d(kernel_size=2, stride=2)def forward(self, x):x = self.conv(self.bn(x))return self.pool(x)class ResDenseNet(nn.Module):def __init__(self, num_classes=1000):super(ResDenseNet, self).__init__()self.stem = nn.Sequential(nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False),nn.BatchNorm2d(64),nn.ReLU(inplace=True),nn.MaxPool2d(kernel_size=3, stride=2, padding=1))self.stage1 = self._make_stage(64, 128, num_layers=4, growth_rate=16)self.stage2 = self._make_stage(128, 256, num_layers=4, growth_rate=16)self.stage3 = self._make_stage(256, 512, num_layers=6, growth_rate=12)self.stage4 = self._make_stage(512, 1024, num_layers=6, growth_rate=12)self.classifier = nn.Linear(1024, num_classes)def _make_stage(self, in_channels, out_channels, num_layers, growth_rate):dense_block = DenseBlock(num_layers, in_channels, growth_rate)transition = TransitionLayer(in_channels + num_layers * growth_rate, out_channels)return nn.Sequential(dense_block, transition)def forward(self, x):x = self.stem(x)x = self.stage1(x)x = self.stage2(x)x = self.stage3(x)x = self.stage4(x)x = torch.mean(x, dim=[2, 3]) # Global Average Poolingreturn self.classifier(x)device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = ResDenseNet().to(device)
model
代码输出：
ResDenseNet((stem): Sequential((0): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)(1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(2): ReLU(inplace=True)(3): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False))(stage1): Sequential((0): DenseBlock((layers): ModuleList((0): Bottleneck((bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(64, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(1): Bottleneck((bn1): BatchNorm2d(80, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(80, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(64, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(2): Bottleneck((bn1): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(96, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(64, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(3): Bottleneck((bn1): BatchNorm2d(112, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(112, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(64, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)))(residual): Conv2d(64, 128, kernel_size=(1, 1), stride=(1, 1), bias=False))(1): TransitionLayer((bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(pool): AvgPool2d(kernel_size=2, stride=2, padding=0)))(stage2): Sequential((0): DenseBlock((layers): ModuleList((0): Bottleneck((bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(64, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(1): Bottleneck((bn1): BatchNorm2d(144, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(144, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(64, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(2): Bottleneck((bn1): BatchNorm2d(160, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(160, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(64, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(3): Bottleneck((bn1): BatchNorm2d(176, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(176, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(64, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)))(residual): Conv2d(128, 192, kernel_size=(1, 1), stride=(1, 1), bias=False))(1): TransitionLayer((bn): BatchNorm2d(192, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv): Conv2d(192, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(pool): AvgPool2d(kernel_size=2, stride=2, padding=0)))(stage3): Sequential((0): DenseBlock((layers): ModuleList((0): Bottleneck((bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(256, 48, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(48, 12, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(1): Bottleneck((bn1): BatchNorm2d(268, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(268, 48, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(48, 12, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(2): Bottleneck((bn1): BatchNorm2d(280, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(280, 48, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(48, 12, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(3): Bottleneck((bn1): BatchNorm2d(292, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(292, 48, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(48, 12, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(4): Bottleneck((bn1): BatchNorm2d(304, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(304, 48, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(48, 12, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(5): Bottleneck((bn1): BatchNorm2d(316, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(316, 48, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(48, 12, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)))(residual): Conv2d(256, 328, kernel_size=(1, 1), stride=(1, 1), bias=False))(1): TransitionLayer((bn): BatchNorm2d(328, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv): Conv2d(328, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(pool): AvgPool2d(kernel_size=2, stride=2, padding=0)))(stage4): Sequential((0): DenseBlock((layers): ModuleList((0): Bottleneck((bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(512, 48, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(48, 12, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(1): Bottleneck((bn1): BatchNorm2d(524, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(524, 48, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(48, 12, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(2): Bottleneck((bn1): BatchNorm2d(536, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(536, 48, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(48, 12, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(3): Bottleneck((bn1): BatchNorm2d(548, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(548, 48, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(48, 12, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(4): Bottleneck((bn1): BatchNorm2d(560, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(560, 48, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(48, 12, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))(5): Bottleneck((bn1): BatchNorm2d(572, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv1): Conv2d(572, 48, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn2): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(48, 12, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)))(residual): Conv2d(512, 584, kernel_size=(1, 1), stride=(1, 1), bias=False))(1): TransitionLayer((bn): BatchNorm2d(584, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv): Conv2d(584, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)(pool): AvgPool2d(kernel_size=2, stride=2, padding=0)))(classifier): Linear(in_features=1024, out_features=1000, bias=True)
)
代码输入：
import torchsummary as summary
summary.summary(model, (3, 224, 224))
代码输出：
----------------------------------------------------------------Layer (type) Output Shape Param #
================================================================Conv2d-1 [-1, 64, 112, 112] 9,408BatchNorm2d-2 [-1, 64, 112, 112] 128ReLU-3 [-1, 64, 112, 112] 0MaxPool2d-4 [-1, 64, 56, 56] 0Conv2d-5 [-1, 128, 56, 56] 8,192BatchNorm2d-6 [-1, 64, 56, 56] 128Conv2d-7 [-1, 64, 56, 56] 4,096BatchNorm2d-8 [-1, 64, 56, 56] 128Conv2d-9 [-1, 16, 56, 56] 9,216Bottleneck-10 [-1, 80, 56, 56] 0BatchNorm2d-11 [-1, 80, 56, 56] 160Conv2d-12 [-1, 64, 56, 56] 5,120BatchNorm2d-13 [-1, 64, 56, 56] 128Conv2d-14 [-1, 16, 56, 56] 9,216Bottleneck-15 [-1, 96, 56, 56] 0BatchNorm2d-16 [-1, 96, 56, 56] 192Conv2d-17 [-1, 64, 56, 56] 6,144BatchNorm2d-18 [-1, 64, 56, 56] 128Conv2d-19 [-1, 16, 56, 56] 9,216Bottleneck-20 [-1, 112, 56, 56] 0BatchNorm2d-21 [-1, 112, 56, 56] 224Conv2d-22 [-1, 64, 56, 56] 7,168BatchNorm2d-23 [-1, 64, 56, 56] 128Conv2d-24 [-1, 16, 56, 56] 9,216Bottleneck-25 [-1, 128, 56, 56] 0DenseBlock-26 [-1, 128, 56, 56] 0BatchNorm2d-27 [-1, 128, 56, 56] 256Conv2d-28 [-1, 128, 56, 56] 16,384AvgPool2d-29 [-1, 128, 28, 28] 0TransitionLayer-30 [-1, 128, 28, 28] 0Conv2d-31 [-1, 192, 28, 28] 24,576BatchNorm2d-32 [-1, 128, 28, 28] 256Conv2d-33 [-1, 64, 28, 28] 8,192BatchNorm2d-34 [-1, 64, 28, 28] 128Conv2d-35 [-1, 16, 28, 28] 9,216Bottleneck-36 [-1, 144, 28, 28] 0BatchNorm2d-37 [-1, 144, 28, 28] 288Conv2d-38 [-1, 64, 28, 28] 9,216BatchNorm2d-39 [-1, 64, 28, 28] 128Conv2d-40 [-1, 16, 28, 28] 9,216Bottleneck-41 [-1, 160, 28, 28] 0BatchNorm2d-42 [-1, 160, 28, 28] 320Conv2d-43 [-1, 64, 28, 28] 10,240BatchNorm2d-44 [-1, 64, 28, 28] 128Conv2d-45 [-1, 16, 28, 28] 9,216Bottleneck-46 [-1, 176, 28, 28] 0BatchNorm2d-47 [-1, 176, 28, 28] 352Conv2d-48 [-1, 64, 28, 28] 11,264BatchNorm2d-49 [-1, 64, 28, 28] 128Conv2d-50 [-1, 16, 28, 28] 9,216Bottleneck-51 [-1, 192, 28, 28] 0DenseBlock-52 [-1, 192, 28, 28] 0BatchNorm2d-53 [-1, 192, 28, 28] 384Conv2d-54 [-1, 256, 28, 28] 49,152AvgPool2d-55 [-1, 256, 14, 14] 0TransitionLayer-56 [-1, 256, 14, 14] 0Conv2d-57 [-1, 328, 14, 14] 83,968BatchNorm2d-58 [-1, 256, 14, 14] 512Conv2d-59 [-1, 48, 14, 14] 12,288BatchNorm2d-60 [-1, 48, 14, 14] 96Conv2d-61 [-1, 12, 14, 14] 5,184Bottleneck-62 [-1, 268, 14, 14] 0BatchNorm2d-63 [-1, 268, 14, 14] 536Conv2d-64 [-1, 48, 14, 14] 12,864BatchNorm2d-65 [-1, 48, 14, 14] 96Conv2d-66 [-1, 12, 14, 14] 5,184Bottleneck-67 [-1, 280, 14, 14] 0BatchNorm2d-68 [-1, 280, 14, 14] 560Conv2d-69 [-1, 48, 14, 14] 13,440BatchNorm2d-70 [-1, 48, 14, 14] 96Conv2d-71 [-1, 12, 14, 14] 5,184Bottleneck-72 [-1, 292, 14, 14] 0BatchNorm2d-73 [-1, 292, 14, 14] 584Conv2d-74 [-1, 48, 14, 14] 14,016BatchNorm2d-75 [-1, 48, 14, 14] 96Conv2d-76 [-1, 12, 14, 14] 5,184Bottleneck-77 [-1, 304, 14, 14] 0BatchNorm2d-78 [-1, 304, 14, 14] 608Conv2d-79 [-1, 48, 14, 14] 14,592BatchNorm2d-80 [-1, 48, 14, 14] 96Conv2d-81 [-1, 12, 14, 14] 5,184Bottleneck-82 [-1, 316, 14, 14] 0BatchNorm2d-83 [-1, 316, 14, 14] 632Conv2d-84 [-1, 48, 14, 14] 15,168BatchNorm2d-85 [-1, 48, 14, 14] 96Conv2d-86 [-1, 12, 14, 14] 5,184Bottleneck-87 [-1, 328, 14, 14] 0DenseBlock-88 [-1, 328, 14, 14] 0BatchNorm2d-89 [-1, 328, 14, 14] 656Conv2d-90 [-1, 512, 14, 14] 167,936AvgPool2d-91 [-1, 512, 7, 7] 0TransitionLayer-92 [-1, 512, 7, 7] 0Conv2d-93 [-1, 584, 7, 7] 299,008BatchNorm2d-94 [-1, 512, 7, 7] 1,024Conv2d-95 [-1, 48, 7, 7] 24,576BatchNorm2d-96 [-1, 48, 7, 7] 96Conv2d-97 [-1, 12, 7, 7] 5,184Bottleneck-98 [-1, 524, 7, 7] 0BatchNorm2d-99 [-1, 524, 7, 7] 1,048Conv2d-100 [-1, 48, 7, 7] 25,152BatchNorm2d-101 [-1, 48, 7, 7] 96Conv2d-102 [-1, 12, 7, 7] 5,184Bottleneck-103 [-1, 536, 7, 7] 0BatchNorm2d-104 [-1, 536, 7, 7] 1,072Conv2d-105 [-1, 48, 7, 7] 25,728BatchNorm2d-106 [-1, 48, 7, 7] 96Conv2d-107 [-1, 12, 7, 7] 5,184Bottleneck-108 [-1, 548, 7, 7] 0BatchNorm2d-109 [-1, 548, 7, 7] 1,096Conv2d-110 [-1, 48, 7, 7] 26,304BatchNorm2d-111 [-1, 48, 7, 7] 96Conv2d-112 [-1, 12, 7, 7] 5,184Bottleneck-113 [-1, 560, 7, 7] 0BatchNorm2d-114 [-1, 560, 7, 7] 1,120Conv2d-115 [-1, 48, 7, 7] 26,880BatchNorm2d-116 [-1, 48, 7, 7] 96Conv2d-117 [-1, 12, 7, 7] 5,184Bottleneck-118 [-1, 572, 7, 7] 0BatchNorm2d-119 [-1, 572, 7, 7] 1,144Conv2d-120 [-1, 48, 7, 7] 27,456BatchNorm2d-121 [-1, 48, 7, 7] 96Conv2d-122 [-1, 12, 7, 7] 5,184Bottleneck-123 [-1, 584, 7, 7] 0DenseBlock-124 [-1, 584, 7, 7] 0BatchNorm2d-125 [-1, 584, 7, 7] 1,168Conv2d-126 [-1, 1024, 7, 7] 598,016AvgPool2d-127 [-1, 1024, 3, 3] 0TransitionLayer-128 [-1, 1024, 3, 3] 0Linear-129 [-1, 1000] 1,025,000
================================================================
Total params: 2,734,104
Trainable params: 2,734,104
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 95.40
Params size (MB): 10.43
Estimated Total Size (MB): 106.41
----------------------------------------------------------------接下来我们简单阅读我们构建的网络：
- 首先我们构建Bottleneck，bottleneck的主要目的是构建denseblock的组成部分，通过两次归一化层以及两次卷积构成
- 随后我们构建Denseblock，并且使用残差连接
- 构建transition层进行池化，最终能够全连接
- 整体网络构建如下：
Input (224x224x3)|| Conv2d (7x7, stride=2)| BatchNorm2d| ReLU| MaxPool2d (3x3, stride=2)v
Stem Layer (64 channels)|v
Stage 1: DenseBlock + TransitionLayer (64 -> 128 channels)| v
Stage 2: DenseBlock + TransitionLayer (128 -> 256 channels)|v
Stage 3: DenseBlock + TransitionLayer (256 -> 512 channels)|v
Stage 4: DenseBlock + TransitionLayer (512 -> 1024 channels)|v
Global Average Pooling (1024x1x1)|v
Fully Connected Layer (1024 -> num_classes)|v
Output (num_classes)二、对上周的乳腺癌识别
import pathlib
data_dir = './data/J3-1-data'
data_dir = pathlib.Path(data_dir)data_path = list(data_dir.glob('*'))
classNames = [path.name for path in data_path]
print(classNames)
代码输出：
['0', '1']from torch.utils.data import DataLoader
from torchvision import datasets, transformstrain_transforms = transforms.Compose([transforms.Resize([224, 224]),transforms.ToTensor(),transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])total_data = datasets.ImageFolder(data_dir, transform=train_transforms)
total_data
代码输出：
Dataset ImageFolderNumber of datapoints: 13403Root location: data\J3-1-dataStandardTransform
Transform: Compose(Resize(size=[224, 224], interpolation=bilinear, max_size=None, antialias=True)ToTensor()Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]))
train_size = int(0.7 * len(total_data))
remain_size = len(total_data) - train_size
train_dataset, remain_dataset = torch.utils.data.random_split(total_data, [train_size, remain_size])
test_size = int(0.6 * len(remain_dataset))
validate_size = len(remain_dataset) - test_size
test_dataset, validate_dataset = torch.utils.data.random_split(remain_dataset, [test_size, validate_size]) #随机分配数据
train_dataset, test_dataset, validate_dataset
代码输出：
(<torch.utils.data.dataset.Subset at 0x2138402dbb0>,<torch.utils.data.dataset.Subset at 0x21383feb590>,<torch.utils.data.dataset.Subset at 0x21383ece690>)
batch_size = 32train_dl = DataLoader(train_dataset, batch_size=batch_size,shuffle=True)test_dl = DataLoader(test_dataset,batch_size = batch_size,shuffle = True
)validate_dl = DataLoader(validate_dataset,batch_size = batch_size,shuffle = False
)for x, y in validate_dl:print("shape of x [N, C, H, W]:", x.shape)print("shape of y:", y.shape, y.dtype)break
代码输出：
shape of x [N, C, H, W]: torch.Size([32, 3, 224, 224])
shape of y: torch.Size([32]) torch.int64def train(dataloader, model, loss_fn, optimizer):size = len(dataloader.dataset)num_batches = len(dataloader)train_loss, train_acc = 0, 0for x, y in dataloader:x, y = x.to(device), y.to(device)pred = model(x)loss = loss_fn(pred, y)#backwardoptimizer.zero_grad()loss.backward()optimizer.step()train_acc += (pred.argmax(1) == y).type(torch.float).sum().item()train_loss += loss.item()train_acc /= sizetrain_loss /= num_batchesreturn train_acc, train_lossdef test(dataloader, model, loss_fn):size = len(dataloader.dataset)num_batches = len(dataloader)test_loss, test_acc = 0, 0for x, y in dataloader:x, y = x.to(device), y.to(device)pred = model(x)loss = loss_fn(pred, y)test_acc += (pred.argmax(1) == y).type(torch.float).sum().item()test_loss += loss.item()test_acc /= sizetest_loss /= num_batchesreturn test_acc, test_loss
训练：
import copy
from torch.optim.lr_scheduler import ReduceLROnPlateauopt = torch.optim.Adam(model.parameters(), lr= 1e-4)
scheduler = ReduceLROnPlateau(opt, mode='min', factor=0.1, patience=5, verbose=True) # 当指标（如损失）连续 5 次没有改善时，将学习率乘以 0.1
loss_fn = nn.CrossEntropyLoss() # 交叉熵epochs = 32train_loss = []
train_acc = []
test_loss = []
test_acc = []best_acc = 0 # 设置一个最佳准确率，作为最佳模型的判别指标for epoch in range(epochs):model.train()epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, opt)model.eval()epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)scheduler.step(epoch_test_loss)if epoch_test_acc > best_acc:best_acc = epoch_test_accbest_model = copy.deepcopy(model)train_acc.append(epoch_train_acc)train_loss.append(epoch_train_loss)test_acc.append(epoch_test_acc)test_loss.append(epoch_test_loss)# 获取当前的学习率lr = opt.state_dict()['param_groups'][0]['lr']template = ('Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%, Test_loss:{:.3f}, Lr:{:.2E}')print(template.format(epoch+1, epoch_train_acc*100, epoch_train_loss, epoch_test_acc*100, epoch_test_loss, lr))# 保存最佳模型到文件中
PATH = './best_model.pth' # 保存的参数文件名
torch.save(best_model.state_dict(), PATH)print('Done')
代码输出：
Epoch: 1, Train_acc:80.7%, Train_loss:0.892, Test_acc:71.2%, Test_loss:1.992, Lr:1.00E-04
Epoch: 2, Train_acc:82.5%, Train_loss:0.409, Test_acc:83.9%, Test_loss:0.393, Lr:1.00E-04
Epoch: 3, Train_acc:83.4%, Train_loss:0.395, Test_acc:82.8%, Test_loss:0.443, Lr:1.00E-04
Epoch: 4, Train_acc:83.8%, Train_loss:0.380, Test_acc:84.1%, Test_loss:0.378, Lr:1.00E-04
Epoch: 5, Train_acc:84.2%, Train_loss:0.375, Test_acc:54.6%, Test_loss:1.337, Lr:1.00E-04
Epoch: 6, Train_acc:84.2%, Train_loss:0.378, Test_acc:84.7%, Test_loss:0.354, Lr:1.00E-04
Epoch: 7, Train_acc:84.7%, Train_loss:0.368, Test_acc:64.4%, Test_loss:0.696, Lr:1.00E-04
Epoch: 8, Train_acc:84.9%, Train_loss:0.360, Test_acc:84.7%, Test_loss:0.493, Lr:1.00E-04
Epoch: 9, Train_acc:85.1%, Train_loss:0.362, Test_acc:73.7%, Test_loss:0.506, Lr:1.00E-04
Epoch:10, Train_acc:85.2%, Train_loss:0.350, Test_acc:77.3%, Test_loss:0.791, Lr:1.00E-04
Epoch:11, Train_acc:85.5%, Train_loss:0.352, Test_acc:53.7%, Test_loss:2.223, Lr:1.00E-04
Epoch:12, Train_acc:85.6%, Train_loss:0.351, Test_acc:84.5%, Test_loss:0.438, Lr:1.00E-05
Epoch:13, Train_acc:86.7%, Train_loss:0.321, Test_acc:87.4%, Test_loss:0.295, Lr:1.00E-05
Epoch:14, Train_acc:86.5%, Train_loss:0.314, Test_acc:87.3%, Test_loss:0.296, Lr:1.00E-05
Epoch:15, Train_acc:87.2%, Train_loss:0.310, Test_acc:87.1%, Test_loss:0.320, Lr:1.00E-05
Epoch:16, Train_acc:87.6%, Train_loss:0.307, Test_acc:87.2%, Test_loss:0.297, Lr:1.00E-05
Epoch:17, Train_acc:87.4%, Train_loss:0.309, Test_acc:88.2%, Test_loss:0.289, Lr:1.00E-05
Epoch:18, Train_acc:87.0%, Train_loss:0.310, Test_acc:87.6%, Test_loss:0.293, Lr:1.00E-05
Epoch:19, Train_acc:87.1%, Train_loss:0.305, Test_acc:88.3%, Test_loss:0.281, Lr:1.00E-05
Epoch:20, Train_acc:87.6%, Train_loss:0.298, Test_acc:87.6%, Test_loss:0.299, Lr:1.00E-05
Epoch:21, Train_acc:87.5%, Train_loss:0.299, Test_acc:87.9%, Test_loss:0.289, Lr:1.00E-05
Epoch:22, Train_acc:87.5%, Train_loss:0.299, Test_acc:88.3%, Test_loss:0.292, Lr:1.00E-05
Epoch:23, Train_acc:88.0%, Train_loss:0.296, Test_acc:86.4%, Test_loss:0.347, Lr:1.00E-05
Epoch:24, Train_acc:87.7%, Train_loss:0.299, Test_acc:88.1%, Test_loss:0.286, Lr:1.00E-05
Epoch:25, Train_acc:87.8%, Train_loss:0.294, Test_acc:86.4%, Test_loss:0.327, Lr:1.00E-06
Epoch:26, Train_acc:87.9%, Train_loss:0.290, Test_acc:87.5%, Test_loss:0.291, Lr:1.00E-06
Epoch:27, Train_acc:88.2%, Train_loss:0.286, Test_acc:88.9%, Test_loss:0.272, Lr:1.00E-06
Epoch:28, Train_acc:88.1%, Train_loss:0.287, Test_acc:88.6%, Test_loss:0.277, Lr:1.00E-06
Epoch:29, Train_acc:88.2%, Train_loss:0.286, Test_acc:89.4%, Test_loss:0.269, Lr:1.00E-06
Epoch:30, Train_acc:88.1%, Train_loss:0.285, Test_acc:89.1%, Test_loss:0.271, Lr:1.00E-06
Epoch:31, Train_acc:88.1%, Train_loss:0.288, Test_acc:88.9%, Test_loss:0.274, Lr:1.00E-06
Epoch:32, Train_acc:87.9%, Train_loss:0.291, Test_acc:89.1%, Test_loss:0.275, Lr:1.00E-06
Done结果上看不如上次的DenseNet121
结果可视化：
import matplotlib.pyplot as plt
epochs_range = range(epochs)plt.figure(figsize=(12, 3))
plt.subplot(1, 2, 1)plt.plot(epochs_range, train_acc, label='Training Accuracy')
plt.plot(epochs_range, test_acc, label='Test Accuracy')
plt.legend(loc='lower right')
plt.title('Training and Validation Accuracy')plt.subplot(1, 2, 2)
plt.plot(epochs_range, train_loss, label='Training Loss')
plt.plot(epochs_range, test_loss, label='Test Loss')
plt.legend(loc='upper right')
plt.title('Training and Validation Loss')
plt.show()
代码输出：
对验证集的准确率：
def validate(dataloader, model):model.eval()size = len(dataloader.dataset)num_batches = len(dataloader)validate_acc = 0for x, y in dataloader:x, y = x.to(device), y.to(device)pred = model(x)validate_acc += (pred.argmax(1) == y).type(torch.float).sum().item()validate_acc /= sizereturn validate_acc# 计算验证集准确率
validate_acc = validate(validate_dl, best_model)
print(f"Validation Accuracy: {validate_acc:.2%}")
代码输出：
Validation Accuracy: 89.37%
达到89.4%
三、总结
这次的结合主要是在和GPT一起完成的，主要是简单的结合，看到很多人说文献中报道过DPN结构，我待会儿也会去看看。
