YOLOv5算法改进--通过yaml文件添加注意力机制【附代码】
以往的算法改进中，网络结构一般直接是用代码写好的，个人感觉这样的方式其实更利于算法的改进，更简单明了。但也有很多算法，比如v5，YOLOR，YOLOV7等等，是用yaml文件进行模型的定义，然后通过pare_model()函数读取模型，如果你需要在模型中添加若干个其他模块，增添自己的想法，通过修改yaml是不易的，但在实际中这又是不可避免的。因此本文章将以YOLOv5 6.0为例，在yaml中添加注意力机制来改进网络。
本文将在yolov5s结构中插入通道注意力机制SENet为例。分别从以下两个方面添加注意力机制：
1.不改变网络深度的改进(直接替换某些层)
2.改变原网络深度(增加层数)
目录
1.不改变网络深度的改进
1.1yaml修改
1.2yolo.py修改
2.改变原网络深度
在添加注意力之前，需要在models/common.py中添加SENet代码。代码如下:
class SE_Block(nn.Module):
def __init__(self, c1, c2):
super().__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1) # 平均池化
self.fc = nn.Sequential(
nn.Linear(c1, c2 // 16, bias=False),
nn.ReLU(inplace=True),
nn.Linear(c2 // 16, c2, bias=False),
nn.Sigmoid()
)
def forward(self, x):
# 添加注意力模块
b, c, _, _ = x.size() # 分别获取batch_size,channel
y = self.avg_pool(x).view(b, c) # y的shape为【batch_size, channels】
y = self.fc(y).view(b, c, 1, 1) # shape为【batch_size, channels, 1, 1】
out = x * y.expand_as(x) # shape 为【batch, channels,feature_w, feature_h】
return out1.不改变网络深度的改进
网络的修改需要修改两个位置：yaml文件、models/yolo.py
1.1yaml修改
首先是打开models/yolov5s.yaml文件，我们在backbone中的SPPF之前增加SENet。增添位置如下，是将backbone中第4个C3模块替换为SE_Block，如下图。
【需要注意的是通道数要匹配，SENet并不改变通道数，由于原C3的输出通道数为1024*0.5=512，所以我们这里的写的是1024，这里的1024是传入到上面我们定义的Class SE_Block(nn.Moudel)中的c2参数，c1参数是由上一层的输出通道数控制的】。
backbone:
# [from, number, module, args]
[[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 conv1(3,32,k=6,s=2,p=2)
[-1, 1, Conv, [128, 3, 2]], # 1-P2/4 conv2(32,64,k=3,s=2,p=1)
[-1, 3, C3, [128]], # C3_1 有Bottleneck
[-1, 1, Conv, [256, 3, 2]], # 3-P3/8 conv3(64,128,k=3,s=2,p=1)
[-1, 6, C3, [256]], # C3_2 Bottleneck重复两次
[-1, 1, Conv, [512, 3, 2]], # 5-P4/16 conv4(128,256,k=3,s=2,p=1)
[-1, 9, C3, [512]], # C3_3 Bottleneck重复三次 输出256通道
[-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 Conv5(256,512,k=3,s=2,p=1)
#[-1, 3, C3, [1024]], # C3_4 Bottleneck重复1次 输出512通道
[-1, 1, SE_Block, [1024]], # 增加通道注意力机制 输出为512通道
[-1, 1, SPPF, [1024, 5]], # 9 每个都是K为5的池化
]1.2yolo.py修改
主要是在下面的代码中，在列表中添加SE_Block，这样可以获得我们要传入的参数。这样就改完了。
if m in [Conv, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, DWConv, MixConv2d, Focus, CrossConv,
BottleneckCSP, C3, C3TR, C3SPP, C3Ghost, SE_Block]:改完以后可以随便训练几轮试试，或者打印一下模型，看看改的对不对。可以看到在我自己的数据集上训练是可以正常进行的，
Epoch gpu_mem box obj cls labels img_size
1/299 1.4G 0.05173 0.03137 0 5 640: 100%|████████████████████████████████████████████████████████████████████████| 180/180 [04:09<00:00, 1.39s/it]
Class Images Labels P R [email protected] [email protected]:.95: 100%|██████████████████████████████████████████████████████████| 10/10 [00:05<00:00, 1.94it/s]
all 80 133 0.416 0.444 0.393 0.118
Epoch gpu_mem box obj cls labels img_size
2/299 1.4G 0.04407 0.02762 0 4 640: 100%|████████████████████████████████████████████████████████████████████████| 180/180 [04:09<00:00, 1.39s/it]
Class Images Labels P R [email protected] [email protected]:.95: 100%|██████████████████████████████████████████████████████████| 10/10 [00:05<00:00, 1.95it/s]
all 80 133 0.655 0.571 0.581 0.21
2.改变原网络深度
比如我要在第一个C3后面加一个SE。yaml的修改如下。接下来稍微麻烦一点了【需要你了解v5的每层结构】，由于我们在backbone中加入了一层，也就是相当于后面的网络与之前相比都往后移动了一层，那么在后面的Concat部分中融合的特征层的索引也会收到影响，因此我们需要的是修改Concat层的from参数。
会改变原来网络中的这些地方：
backbone:
# [from, number, module, args]
[[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 conv1(3,32,k=6,s=2,p=2)
[-1, 1, Conv, [128, 3, 2]], # 1-P2/4 conv2(32,64,k=3,s=2,p=1)
[-1, 3, C3, [128]], # C3_1 有Bottleneck
[-1, 1, SE_Block, [128]], # 增加通道注意力机制 输出为512通道
[-1, 1, Conv, [256, 3, 2]], # 3-P3/8 conv3(64,128,k=3,s=2,p=1)
[-1, 6, C3, [256]], # C3_2 Bottleneck重复两次
[-1, 1, Conv, [512, 3, 2]], # 5-P4/16 conv4(128,256,k=3,s=2,p=1)
[-1, 9, C3, [512]], # C3_3 Bottleneck重复三次 输出256通道
[-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 Conv5(256,512,k=3,s=2,p=1)
[-1, 3, C3, [1024]], # C3_4 Bottleneck重复1次 输出512通道
[-1, 1, SPPF, [1024, 5]], # 9 每个都是K为5的池化
]可以看到实际就是每个Concat也后面移动一层，因此yaml修改为一下。最终的Detect的from也需要修改。
head:
[[-1, 1, Conv, [512, 1, 1]], # conv1(512,256,1,1)
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[[-1, 7], 1, Concat, [1]], # cat backbone P4 将C3_3与SPPF出来后的上采样拼接 拼接后的通道为512
[-1, 3, C3, [512, False]], # 13 conv(256,256,k=1,s=1) 没有残差边
[-1, 1, Conv, [256, 1, 1]], # conv2(256,128,1,1)
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[[-1, 5], 1, Concat, [1]], # cat backbone P3 与C3_2拼接，输出256通道
[-1, 3, C3, [256, False]], # 17 (P3/8-small) conv3(128,128,1,1)
[-1, 1, Conv, [256, 3, 2]],# conv4(128,128,3,2,1)
[[-1, 15], 1, Concat, [1]], # cat head P4 拼接后256通道
[-1, 3, C3, [512, False]], # 20 (P4/16-medium) conv5(256,256,1,1)
[-1, 1, Conv, [512, 3, 2]],# conv6(256,256,3,2,1)
[[-1, 11], 1, Concat, [1]], # cat head P5 拼接后是512
[-1, 3, C3, [1024, False]], # 23 (P5/32-large)
[[18, 21, 24], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5)
]同理，也需要在models/yolo.py中修改如下，增加SE_Block：
if m in [Conv, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, DWConv, MixConv2d, Focus, CrossConv,
BottleneckCSP, C3, C3TR, C3SPP, C3Ghost, SE_Block]:也可以训练一个epoch看看，或者打印一下网络。这样就改完啦。
Starting training for 300 epochs...
Epoch gpu_mem box obj cls labels img_size
0/299 0.419G 0.09617 0.03019 0 13 640: 12%|████████▉ | 22/180 [00:15<01:05, 2.40it/s]
Model(
(model): Sequential(
(0): Conv(
(conv): Conv2d(3, 32, kernel_size=(6, 6), stride=(2, 2), padding=(2, 2), bias=False)
(bn): BatchNorm2d(32, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(1): Conv(
(conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(2): C3(
(cv1): Conv(
(conv): Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(32, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(32, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv3): Conv(
(conv): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): Sequential(
(0): Bottleneck(
(cv1): Conv(
(conv): Conv2d(32, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(32, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(32, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
)
)
)
(3): SE_Block(
(avg_pool): AdaptiveAvgPool2d(output_size=1)
(fc): Sequential(
(0): Linear(in_features=64, out_features=4, bias=False)
(1): ReLU(inplace=True)
(2): Linear(in_features=4, out_features=64, bias=False)
(3): Sigmoid()
)
)
(4): Conv(
(conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(5): C3(
(cv1): Conv(
(conv): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv3): Conv(
(conv): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): Sequential(
(0): Bottleneck(
(cv1): Conv(
(conv): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
)
(1): Bottleneck(
(cv1): Conv(
(conv): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
)
)
)
(6): Conv(
(conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(7): C3(
(cv1): Conv(
(conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv3): Conv(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): Sequential(
(0): Bottleneck(
(cv1): Conv(
(conv): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
)
(1): Bottleneck(
(cv1): Conv(
(conv): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
)
(2): Bottleneck(
(cv1): Conv(
(conv): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
)
)
)
(8): Conv(
(conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(512, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(9): C3(
(cv1): Conv(
(conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv3): Conv(
(conv): Conv2d(512, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(512, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): Sequential(
(0): Bottleneck(
(cv1): Conv(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
)
)
)
(10): SPPF(
(cv1): Conv(
(conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(512, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): MaxPool2d(kernel_size=5, stride=1, padding=2, dilation=1, ceil_mode=False)
)
(11): Conv(
(conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(12): Upsample(scale_factor=2.0, mode=nearest)
(13): Concat()
(14): C3(
(cv1): Conv(
(conv): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv3): Conv(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): Sequential(
(0): Bottleneck(
(cv1): Conv(
(conv): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
)
)
)
(15): Conv(
(conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(16): Upsample(scale_factor=2.0, mode=nearest)
(17): Concat()
(18): C3(
(cv1): Conv(
(conv): Conv2d(256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv3): Conv(
(conv): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): Sequential(
(0): Bottleneck(
(cv1): Conv(
(conv): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
)
)
)
(19): Conv(
(conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(20): Concat()
(21): C3(
(cv1): Conv(
(conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv3): Conv(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): Sequential(
(0): Bottleneck(
(cv1): Conv(
(conv): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
)
)
)
(22): Conv(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(23): Concat()
(24): C3(
(cv1): Conv(
(conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv3): Conv(
(conv): Conv2d(512, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(512, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): Sequential(
(0): Bottleneck(
(cv1): Conv(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU()
)
)
)
)
(25): Detect(
(m): ModuleList(
(0): Conv2d(128, 255, kernel_size=(1, 1), stride=(1, 1))
(1): Conv2d(256, 255, kernel_size=(1, 1), stride=(1, 1))
(2): Conv2d(512, 255, kernel_size=(1, 1), stride=(1, 1))
)
)
)
)
不论哪种网络的修改，都需要注意的一点是通道维度的对应，如果报错了，例如mat1和mat2无法相乘，或者shape匹配问题，都是你修改后的网络channels不匹配导致的，要细心检查，在增删网络的时候，要注意yaml中from列表参数的影响，因为你修改了一层，将会影响后面的后面层。