https://github.com/xiaqing10/Hisi_YoLoV5

海思nnie跑yolov5
https://github.com/xiaqing10/Hisi_YoLoV5

Last synced: 12 days ago
JSON representation

海思nnie跑yolov5

• Host: GitHub
• URL: https://github.com/xiaqing10/Hisi_YoLoV5
• Owner: xiaqing10
• Created: 2021-09-13T05:18:44.000Z (almost 3 years ago)
• Default Branch: main
• Last Pushed: 2022-07-15T05:54:39.000Z (almost 2 years ago)
• Last Synced: 2024-02-29T09:33:29.597Z (4 months ago)
• Size: 23.5 MB
• Stars: 23
• Watchers: 1
• Forks: 8
• Open Issues: 4
• Metadata Files:
  • Readme: README.md

Lists

• awesome-yolo-object-detection - xiaqing10/Hisi_YoLoV5

README

        
## 鉴于经常还是有人问我尽管有模型，但是还是不知道怎么修改工程，我传上之前的整个测试工程，仅供参考.

# Hisi_YoLoV5

海思nnie跑yolov5

## 模型修改。

### 参考官方的yolov5 4.0版本

+ git clone -b v4.0 https://github.com/ultralytics/yolov5.git/ 

### 修改处0, models/yolov5s.py  增加如下:

+    删除 ~~[[-1, 1, Focus, [64, 3]],  # 0-P1/2 ~~

+   添加 [[-1, 1, Conv, [64, 3, 2]],  # 0-P1/2

### 修改处1, utils/activations.py  增加如下:

    class ReLU(nn.Module): 

        @staticmethod

            def forward(x):

                return nn.ReLU(x)

### 修改处2，models/common.py 35行修改激活函数

+    ~~self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())~~

+    添加 self.act = nn.ReLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())

### 修改处3, models/comon.py 98行修改maxPool ,修改ceil_mode方式

+    self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2, ceil_mode = True) for x in k])

### 修改处4 modes/yolov5s_

+   ~~[-1, 1, nn.Upsample, [None, 2, 'nearest']],~~

+   添加 [-1, 1, nn.ConvTranspose2d, [256,256, 2, 2]],

### 修改处5, models/yolov5s.yaml

+   ~~[-1, 1, nn.Upsample, [None, 2, 'nearest']],~~

+   添加 [-1, 1, nn.ConvTranspose2d, [128 ,128, 2, 2]],

### 转换ONNX修改： models/export.py , 修改了opset_version

+    torch.onnx.export(model, img, f, verbose=False, opset_version=10, input_names=['images']...

### 参考 https://github.com/Wulingtian/yolov5_caffe

+    将ONNX 转为caffe模型。

### 最后caffe模型中的premute和reshape层，nnie不支持，去掉premute层(一共有三层)reshape改为需要的输出。