diff --git a/src/Yolov4_Vehicle_inspection/nets/CSPdarknet.py b/src/Yolov4_Vehicle_inspection/nets/CSPdarknet.py index b6f7965bfc..d9ad967190 100644 --- a/src/Yolov4_Vehicle_inspection/nets/CSPdarknet.py +++ b/src/Yolov4_Vehicle_inspection/nets/CSPdarknet.py @@ -13,6 +13,7 @@ def __init__(self): super(Mish, self).__init__() def forward(self, x): + # 激活函数计算:x * tanh(ln(1 + e^x)),平滑非单调特性提升梯度流动 return x * torch.tanh(F.softplus(x)) #---------------------------------------------------# @@ -23,7 +24,9 @@ class BasicConv(nn.Module): def __init__(self, in_channels, out_channels, kernel_size, stride=1): super(BasicConv, self).__init__() + # 卷积层采用same padding(kernel_size//2),无偏置(BN层替代偏置作用) self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride, kernel_size//2, bias=False) + # BN层加速训练收敛,降低过拟合风险 self.bn = nn.BatchNorm2d(out_channels) self.activation = Mish() @@ -43,13 +46,14 @@ def __init__(self, channels, hidden_channels=None): if hidden_channels is None: hidden_channels = channels - + # 残差块内部:1x1卷积降维 -> 3x3卷积恢复维度,减少计算量同时保留特征 self.block = nn.Sequential( BasicConv(channels, hidden_channels, 1), BasicConv(hidden_channels, channels, 3) ) def forward(self, x): + # 残差连接:原始输入 + 卷积输出,避免梯度消失 return x + self.block(x) #--------------------------------------------------------------------# @@ -76,12 +80,12 @@ def __init__(self, in_channels, out_channels, num_blocks, first): #----------------------------------------------------------------# # 主干部分会对num_blocks进行循环,循环内部是残差结构。 #----------------------------------------------------------------# - self.split_conv1 = BasicConv(out_channels, out_channels, 1) + self.split_conv1 = BasicConv(out_channels, out_channels, 1) self.blocks_conv = nn.Sequential( Resblock(channels=out_channels, hidden_channels=out_channels//2), BasicConv(out_channels, out_channels, 1) ) - + # 拼接后通道数翻倍,1x1卷积整合通道维度 self.concat_conv = BasicConv(out_channels*2, out_channels, 1) else: #--------------------------------------------------------------------------# @@ -93,6 +97,7 @@ def __init__(self, in_channels, out_channels, num_blocks, first): # 主干部分会对num_blocks进行循环,循环内部是残差结构。 #----------------------------------------------------------------# self.split_conv1 = BasicConv(out_channels, out_channels//2, 1) + # 批量构建num_blocks个残差块,形成主干特征提取路径 self.blocks_conv = nn.Sequential( *[Resblock(out_channels//2) for _ in range(num_blocks)], BasicConv(out_channels//2, out_channels//2, 1) @@ -101,16 +106,20 @@ def __init__(self, in_channels, out_channels, num_blocks, first): self.concat_conv = BasicConv(out_channels, out_channels, 1) def forward(self, x): + # 下采样:特征图尺寸减半,通道数翻倍 x = self.downsample_conv(x) + # 大残差边路径:直接卷积,不经过残差堆叠 x0 = self.split_conv0(x) + # 主干路径:经过残差块堆叠提取特征 x1 = self.split_conv1(x) x1 = self.blocks_conv(x1) #------------------------------------# # 将大残差边再堆叠回来 #------------------------------------# + # dim=1按通道维度拼接,x1(主干) + x0(大残差边) x = torch.cat([x1, x0], dim=1) #------------------------------------# # 最后对通道数进行整合 @@ -130,6 +139,7 @@ def __init__(self, layers): self.inplanes = 32 # 416,416,3 -> 416,416,32 self.conv1 = BasicConv(3, self.inplanes, kernel_size=3, stride=1) + # 各阶段输出通道数,对应5次下采样的通道维度 self.feature_channels = [64, 128, 256, 512, 1024] self.stages = nn.ModuleList([ @@ -137,15 +147,16 @@ def __init__(self, layers): Resblock_body(self.inplanes, self.feature_channels[0], layers[0], first=True), # 208,208,64 -> 104,104,128 Resblock_body(self.feature_channels[0], self.feature_channels[1], layers[1], first=False), - # 104,104,128 -> 52,52,256 + # 104,104,128 -> 52,52,256 (输出特征层1:检测小目标) Resblock_body(self.feature_channels[1], self.feature_channels[2], layers[2], first=False), - # 52,52,256 -> 26,26,512 + # 52,52,256 -> 26,26,512 (输出特征层2:检测中目标) Resblock_body(self.feature_channels[2], self.feature_channels[3], layers[3], first=False), - # 26,26,512 -> 13,13,1024 + # 26,26,512 -> 13,13,1024(输出特征层3:检测大目标) Resblock_body(self.feature_channels[3], self.feature_channels[4], layers[4], first=False) ]) self.num_features = 1 + # 权重初始化:卷积层使用He正态初始化,BN层默认初始化 for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels @@ -156,21 +167,29 @@ def __init__(self, layers): def forward(self, x): + # 初始卷积层:提取基础特征 x = self.conv1(x) + # 前两个阶段:基础特征下采样 x = self.stages[0](x) x = self.stages[1](x) + # 第三阶段输出:52x52x256(小目标检测特征层) out3 = self.stages[2](x) + # 第四阶段输出:26x26x512(中目标检测特征层) out4 = self.stages[3](out3) + # 第五阶段输出:13x13x1024(大目标检测特征层) out5 = self.stages[4](out4) + # 返回三个不同尺度的特征层,适配YOLO多尺度检测 return out3, out4, out5 def darknet53(pretrained, **kwargs): + # layers=[1,2,8,8,4] 对应CSPDarkNet53的残差块数量配置 model = CSPDarkNet([1, 2, 8, 8, 4]) if pretrained: if isinstance(pretrained, str): + # 加载预训练权重,需保证权重文件与模型结构匹配 model.load_state_dict(torch.load(pretrained)) else: raise Exception("darknet request a pretrained path. got [{}]".format(pretrained)) - return model + return model \ No newline at end of file