bash
复制
1pip uninstall ultralytics
2pip uninstall ultralytics-thop
3 

Python
复制
1import numpy as np
2import torch
3from torch import nn
4from torch.nn import init
5from torch.nn.parameter import Parameter
6 
7class ShuffleAttention(nn.Module):
8    def __init__(self, channel=512, out_channel=512, reduction=16, G=8):
9        super().__init__()
10        self.G = G
11        self.channel = channel
12        self.avg_pool = nn.AdaptiveAvgPool2d(1)
13        self.gn = nn.GroupNorm(channel // (2 * G), channel // (2 * G))
14        self.cweight = Parameter(torch.zeros(1, channel // (2 * G), 1, 1))
15        self.cbias = Parameter(torch.ones(1, channel // (2 * G), 1, 1))
16        self.sweight = Parameter(torch.zeros(1, channel // (2 * G), 1, 1))
17        self.sbias = Parameter(torch.ones(1, channel // (2 * G), 1, 1))
18        self.sigmoid = nn.Sigmoid()
19    def init_weights(self):
20        for m in self.modules():
21            if isinstance(m, nn.Conv2d):
22                init.kaiming_normal_(m.weight, mode='fan_out')
23                if m.bias is not None:
24                    init.constant_(m.bias, 0)
25            elif isinstance(m, nn.BatchNorm2d):
26                init.constant_(m.weight, 1)
27                init.constant_(m.bias, 0)
28            elif isinstance(m, nn.Linear):
29                init.normal_(m.weight, std=0.001)
30                if m.bias is not None:
31                    init.constant_(m.bias, 0)
32    @staticmethod
33    def channel_shuffle(x, groups):
34        b, c, h, w = x.shape
35        x = x.reshape(b, groups, -1, h, w)
36        x = x.permute(0, 2, 1, 3, 4)
37        # flatten
38        x = x.reshape(b, -1, h, w)
39        return x
40    def forward(self, x):
41        b, c, h, w = x.size()
42        # group into subfeatures
43        x = x.view(b * self.G, -1, h, w)  # bs*G,c//G,h,w
44        # channel_split
45        x_0, x_1 = x.chunk(2, dim=1)  # bs*G,c//(2*G),h,w
46        # channel attention
47        x_channel = self.avg_pool(x_0)  # bs*G,c//(2*G),1,1
48        x_channel = self.cweight * x_channel + self.cbias  # bs*G,c//(2*G),1,1
49        x_channel = x_0 * self.sigmoid(x_channel)
50        # spatial attention
51        x_spatial = self.gn(x_1)  # bs*G,c//(2*G),h,w
52        x_spatial = self.sweight * x_spatial + self.sbias  # bs*G,c//(2*G),h,w
53        x_spatial = x_1 * self.sigmoid(x_spatial)  # bs*G,c//(2*G),h,w
54        # concatenate along channel axis
55        out = torch.cat([x_channel, x_spatial], dim=1)  # bs*G,c//G,h,w
56        out = out.contiguous().view(b, -1, h, w)
57        # channel shuffle
58        out = self.channel_shuffle(out, 2)
59        return out
60 

Python
复制
1from .SA import ShuffleAttention
2 

Python
复制
1将 "ShuffleAttention" 加到 __all__中
2 

Python
复制
1from ultralytics.nn.modules import ShuffleAttention
2 

Python
复制
1elif m is torch.nn.BatchNorm2d:
2	args = [ch[f]]
3 

Python
复制
1elif m is ShuffleAttention: # add shuffle attention
2	c1, c2 = ch[f], args[0]
3	if c2 != nc:  # if c2 not equal to number of classes (i.e. for Classify() output)  
4		c2 = make_divisible(min(c2, max_channels) * width, 8)
5	args = [c1, c2, *args[1:]]
6 

yaml
复制
1# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
2# Ultralytics YOLOv8 object detection model with P3/8 - P5/32 outputs
3# Model docs: https://docs.ultralytics.com/models/yolov8
4# Task docs: https://docs.ultralytics.com/tasks/detect
5 
6# Parameters
7nc: 80 # number of classes
8scales: # model compound scaling constants, i.e. 'model=yolov8n.yaml' will call yolov8.yaml with scale 'n'
9  # [depth, width, max_channels]
10  n: [0.33, 0.25, 1024] # YOLOv8n summary: 129 layers, 3157200 parameters, 3157184 gradients, 8.9 GFLOPS
11  s: [0.33, 0.50, 1024] # YOLOv8s summary: 129 layers, 11166560 parameters, 11166544 gradients, 28.8 GFLOPS
12  m: [0.67, 0.75, 768] # YOLOv8m summary: 169 layers, 25902640 parameters, 25902624 gradients, 79.3 GFLOPS
13  l: [1.00, 1.00, 512] # YOLOv8l summary: 209 layers, 43691520 parameters, 43691504 gradients, 165.7 GFLOPS
14  x: [1.00, 1.25, 512] # YOLOv8x summary: 209 layers, 68229648 parameters, 68229632 gradients, 258.5 GFLOPS
15  
16# YOLOv8.0n backbone
17backbone:
18  # [from, repeats, module, args]
19  - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
20  - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
21  - [-1, 3, C2f, [128, True]]
22  - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
23  - [-1, 6, C2f, [256, True]]
24  - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
25  - [-1, 6, C2f, [512, True]]
26  - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
27  - [-1, 3, C2f, [1024, True]]
28  - [-1, 3, ShuffleAttention, [1024]]
29  - [-1, 1, SPPF, [1024, 5]] # 9
30 
31# YOLOv8.0n head
32head:
33  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
34  - [[-1, 6], 1, Concat, [1]] # cat backbone P4
35  - [-1, 3, C2f, [512]] # 12
36  
37  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
38  - [[-1, 4], 1, Concat, [1]] # cat backbone P3
39  - [-1, 3, C2f, [256]] # 15 (P3/8-small)
40 
41  - [-1, 1, ShuffleAttention, [256, 16, 8]]
42  - [-1, 1, Conv, [256, 3, 2]]
43  - [[-1, 12], 1, Concat, [1]] # cat head P4
44  - [-1, 3, C2f, [512]] # 18 (P4/16-medium)
45  
46  - [-1, 1, ShuffleAttention, [512, 16, 8]]
47  - [-1, 1, Conv, [512, 3, 2]]
48  - [[-1, 9], 1, Concat, [1]] # cat head P5
49  - [-1, 3, C2f, [1024]] # 21 (P5/32-large)
50 
51  - [[15, 18, 21], 1, Detect, [nc]] # Detect(P3, P4, P5)
52 

ln
复制
1import torch
2# Load a model
3model = YOLO("ultralytics/cfg/models/v8/YOLOv8-SA.yaml")  # build a new model from scratch
4ckpt = torch.load("pre_weight/yolov8n.pt", map_location="cpu")
5state = ckpt["model"].float().state_dict()
6model.model.load_state_dict(state, strict=False)    # 只匹配同名层
7