Skip to content
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
33 changes: 26 additions & 7 deletions src/Yolov4_Vehicle_inspection/nets/CSPdarknet.py
Original file line number Diff line number Diff line change
Expand Up @@ -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))

#---------------------------------------------------#
Expand All @@ -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()

Expand All @@ -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)

#--------------------------------------------------------------------#
Expand All @@ -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:
#--------------------------------------------------------------------------#
Expand All @@ -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)
Expand All @@ -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
Expand All @@ -130,22 +139,24 @@ 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([
# 416,416,32 -> 208,208,64
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
Expand All @@ -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
Loading