文章目录
一、设置网络结构为mobilenet-V2
首先，需要在models/common.py里，实现MobileNetv2的 bottleneck 和 Pwconv。
1、Mobilenetv2的bottleneck: InvertedResidual
#mobilenet Bottleneck InvertedResidual
class BottleneckMOB(nn.Module):
#c1:inp c2:oup s:stride expand_ratio:t
def __init__(self, c1, c2, s, expand_ratio):
super(BottleneckMOB, self).__init__()
self.s = s
hidden_dim = round(c1 * expand_ratio)
self.use_res_connect = self.s == 1 and c1 == c2
if expand_ratio == 1:
self.conv = nn.Sequential(
# dw
nn.Conv2d(hidden_dim, hidden_dim, 3, s, 1, groups=hidden_dim, bias=False),
nn.BatchNorm2d(hidden_dim),
nn.ReLU6(inplace=True),
# pw-linear
nn.Conv2d(hidden_dim, c2, 1, 1, 0, bias=False),
nn.BatchNorm2d(c2),
)
else:
self.conv = nn.Sequential(
# pw
nn.Conv2d(c1, hidden_dim, 1, 1, 0, bias=False),
nn.BatchNorm2d(hidden_dim),
nn.ReLU6(inplace=True),
# dw
nn.Conv2d(hidden_dim, hidden_dim, 3, s, 1, groups=hidden_dim, bias=False),
nn.BatchNorm2d(hidden_dim),
nn.ReLU6(inplace=True),
# pw-linear
nn.Conv2d(hidden_dim, c2, 1, 1, 0, bias=False),
nn.BatchNorm2d(c2),
)
def forward(self, x):
if self.use_res_connect:
return x + self.conv(x)
else:
return self.conv(x)
2、Pointwise Convolution
class PW_Conv(nn.Module):
def __init__(self, c1, c2): # ch_in, ch_out
super(PW_Conv, self).__init__()
self.conv = nn.Conv2d(c1, c2, 1, 1, 0, bias=False)
self.bn = nn.BatchNorm2d(c2)
self.act = nn.ReLU6(inplace=True)
def forward(self, x):
return self.act(self.bn(self.conv(x)))
接着需要在yolov5的读取模型配置文件的代码（models/yolo.py的parse_model函数）进行修改，使得能够调用到上面的模块，只需修改下面这部分代码：
n = max(round(n * gd), 1) if n > 1 else n # depth gain
if m in [nn.Conv2d, Conv, Bottleneck, SPP, DWConv, MixConv2d, Focus, CrossConv, BottleneckCSP, C3, PW_Conv, BottleneckMOB]:
c1, c2 = ch[f], args[0]
并且需要在import引用处加入PW_Conv,BottleneckMOB这两个模块：
from models.common import Conv, Bottleneck,SPP, DWConv, Focus, BottleneckCSP, Concat, NMS, autoShape, PW_Conv,BottleneckMOB
然后就是搭建我们的模型配置文件，我在yolov5s.yaml的基础上进行修改，将yolov5s的backbone替换成mobilenetv2，重新建立了一个模型配置文件yolov5-mobilenetV2.yaml：
# parameters
nc: 1 # number of classes
depth_multiple: 0.33 # model depth multiple
width_multiple: 0.50 # layer channel multiple
# anchors
anchors:
- [116,90, 156,198, 373,326] # P5/32
- [30,61, 62,45, 59,119] # P4/16
- [10,13, 16,30, 33,23] # P3/8
# YOLOv5 backbone: mobilenet v2
backbone:
# [from, number, module, args]
[[-1, 1, nn.Conv2d, [32, 3, 2]], # 0-P1/2 oup, k, s 640
[-1, 1, BottleneckMOB, [16, 1, 1]], # 1-P2/4 oup, s, t 320
[-1, 2, BottleneckMOB, [24, 2, 6]], # 320
[-1, 1, PW_Conv, [256]], #4 output p3 160
[-1, 3, BottleneckMOB, [32, 2, 6]], # 3-P3/8 160
[-1, 4, BottleneckMOB, [64, 1, 6]], # 5 80
[-1, 1, PW_Conv, [512]], #7 output p4 6 40
[-1, 3, BottleneckMOB, [96, 2, 6]], # 7 80
[-1, 3, BottleneckMOB, [160, 1, 6,]], # 40
[-1, 1, BottleneckMOB, [320, 1, 6,]], # 40
[-1, 1, nn.Conv2d, [1280, 1, 1]], # 40
[-1, 1, SPP, [1024, [5, 9, 13]]], #11 # 40
]
# YOLOv5 head
head:
[[-1, 3, BottleneckCSP, [1024, False]], # 12 40
[-1, 1, Conv, [512, 1, 1]], # 40
[-1, 1, nn.Upsample, [None, 2, 'nearest']], # 40
[[-1, 6], 1, Concat, [1]], # cat backbone P4-7 # 80
[-1, 3, BottleneckCSP, [512, False]], # 16 # 80
[-1, 1, Conv, [256, 1, 1]], # 80
[-1, 1, nn.Upsample, [None, 2, 'nearest']], # 160
[[-1, 3], 1, Concat, [1]], # cat backbone P3-4 160
[-1, 3, BottleneckCSP, [256, False]], # 160
[-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 21 (P3/8-small) # 160
[-2, 1, Conv, [256, 3, 2]], # 160
[[-1, 17], 1, Concat, [1]], # cat head P4 # 160
[-1, 3, BottleneckCSP, [512, False]], # 160
[-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 25 (P4/16-medium) # 160
[-2, 1, Conv, [512, 3, 2]], # 160
[[-1, 13], 1, Concat, [1]], # cat head P5-13 # 160
[-1, 3, BottleneckCSP, [1024, False]], # 160
[-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 29 (P5/32-large) 160
[[21, 25, 29], 1, Detect, [nc, anchors]], # Detect(P5, P4, P3) nc:number class, na:number of anchors
]
到这我们就实现了将yolov5的backbone替换成了mobilenetv2。在使用时只需要将网络结构配置参数—cfg修改成 –cfg yolov5-mobilenet.yaml。
训练指令：
python train.py --data coco.yaml --cfg yolov5-mobilenet.yaml--weights '' --batch-size 64
二、添加注意力模块
配置文件yolov5x_se.yaml
# parameters
nc: 15 # number of classes
depth_multiple: 1 # model depth multiple
width_multiple: 1 # layer channel multiple
# anchors
anchors:
- [10, 13, 16, 30, 33, 23] # P3/8
- [30, 61, 62, 45, 59, 119] # P4/16
- [116, 90, 156, 198, 373, 326] # P5/32
# YOLOv5 backbone
backbone:
# [from, number, module, args]
[
[-1, 1, Focus, [64, 3]], # 0-P1/2 #1
[-1, 1, Conv, [128, 3, 2]], # 1-P2/4 #2
[-1, 3, C3, [128]], #3
[-1, 1, Conv, [256, 3, 2]], # 3-P3/8 #4
[-1, 9, C3, [256]], #5
[-1, 1, Conv, [512, 3, 2]], # 5-P4/16 #6
[-1, 9, C3, [512]], #7
[-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 #8
[-1, 1, SPP, [1024, [5, 9, 13]]], #9
[-1, 3, C3, [1024, False]], # 9 #10
[-1, 1, SELayer, [1024, 4]], #10
]
# YOLOv5 head
head: [
[-1, 1, Conv, [512, 1, 1]], #11 /32
[-1, 1, nn.Upsample, [None, 2, "nearest"]], #12 /16
[[-1, 6], 1, Concat, [1]], # cat backbone P4 /16 #13
[-1, 3, C3, [512, False]], # 13 / 16 #14
[-1, 1, Conv, [256, 1, 1]], #15 /16
[-1, 1, nn.Upsample, [None, 2, "nearest"]], #16 /8
[[-1, 4], 1, Concat, [1]], # cat backbone P3 /8 #17
[-1, 3, C3, [256, False]], # 17 (P3/8-small) /8 #18
[-1, 1, Conv, [256, 3, 2]], #19 /16
[[-1, 6], 1, Concat, [1]], # cat head P4 #20
[-1, 3, C3, [512, False]], # 20 (P4/16-medium) #21
[-1, 1, Conv, [512, 3, 2]], #22 /32
[[-1, 8], 1, Concat, [1]], # cat head P5 #23
[-1, 3, C3, [1024, False]], # 23 (P5/32-large) #24
[[18, 21, 24], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5)
]
在backbone最后一层添加了SELayer，这个类我已经在common.py中添加进来：
class SELayer(nn.Module):
def __init__(self, c1, r=16):
super(SELayer, self).__init__()
self.avgpool = nn.AdaptiveAvgPool2d(1)
self.l1 = nn.Linear(c1, c1//r, bias=False)
self.relu = nn.ReLU(inplace=True)
self.l2 = nn.Linear(c1//r, c1, bias=False)
self.sig = nn.Sigmoid()
def forward(self, x):
b, c, _, _ = x.size()
y = self.avgpool(x).view(b, c)
y = self.l1(y)
y = self.relu(y)
y = self.l2(y)
y = self.sig(y)
y = y.view(b, c, 1, 1)
return x * y.expand_as(x)
还需要在yolo.py中添加这个改动：
for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']):
m = eval(m) if isinstance(m, str) else m # eval strings
for j, a in enumerate(args):
try:
args[j] = eval(a) if isinstance(a, str) else a # eval strings
except:
pass
n = max(round(n * gd), 1) if n > 1 else n # depth gain
if m in [Conv, GhostConv, Bottleneck, GhostBottleneck, SPP,
DWConv, MixConv2d, Focus, CrossConv, BottleneckCSP,
C3]:
c1, c2 = ch[f], args[0]
if c2 != no: # if not output
c2 = make_divisible(c2 * gw, 8)
args = [c1, c2, *args[1:]]
if m in [BottleneckCSP, C3]:
args.insert(2, n) # number of repeats
n = 1
elif m is nn.BatchNorm2d:
args = [ch[f]]
elif m is Concat:
c2 = sum([ch[x] for x in f])
elif m is Detect:
args.append([ch[x] for x in f])
if isinstance(args[1], int): # number of anchors
args[1] = [list(range(args[1] * 2))] * len(f)
elif m is Contract:
c2 = ch[f] * args[0] ** 2
elif m is Expand:
c2 = ch[f] // args[0] ** 2
elif m is SELayer: # 这里是修改的部分
channel, re = args[0], args[1]
channel = make_divisible(channel * gw, 8) if channel != no else channel
args = [channel, re]
else:
c2 = ch[f]