YOLOv8-Pose NCNN安卓部署
前言
YOLOv8-Pose NCNN安卓部署
目前的帧率可以稳定在30帧左右，下面是这个项目的github地址：https://github.com/gaoxumustwin/ncnn-android-yolov8-pose
介绍
在做YOLOv8-Pose NCNN安卓部署的时候，在github上发现了https://github.com/eecn/ncnn-android-yolov8-pose已经有大佬实现了这一操，并且也给出了它是如何导出ncnn模型的；
该作者是使用YOLOv8的官方给出的NCNN模型导出的方式，我按照这个方式导出ncnn模型，并观察其网络结构如下：
这个网络结构有一些错综复杂，感觉是有大量的计算图追踪，但是其安卓部署的时候速度在20帧左右也还可以接受，并且YOLOv8的官方给出了NCNN推理的python代码
由于之前看到了三木君大佬，在23年的时候在知乎发布的一篇文章,只导出模型的Backbone + Neck,这种思路在玩瑞芯微设备的时候多次接触，三木君大佬也给出了原因：
- 只导出模型，不导出后处理方便对模型进行公平的测速，
MMYOLO里面的算法多种多样，而且大多是Backbone+Neck结构，这样测速才能较为公平体现模型的速度差异。 - 有些模型需要部署到嵌入式设备，导出后处理的话相当与会有海量的加减乘除要计算，由于模型应用会通过置信度过滤很多检测结果，会导致大多计算都是多余的。
- 有些推理后端对后处理算子支持很差，可能会无法转换到对应后端的模型，或者转换后的模型无法充分利用 BPU/TPU/NPU/GPU 加速等等问题存在。
- 很多要部署的模型都会用
c/c++重写后处理加速，后处理代码可以多线程计算，异步计算效率会更高。
同时在CSDN上发现了有佬根据上面的思路编写了NCNN代码并且给出了NCNN代码的github地址
此时只需要参照前面的yolov8安卓部署实现把这个只导出模型的Backbone + Neck的YOLOv8-Pose NCNN代码进行简单改写就可以实现
环境
我使用的ultralytics版本如下：
sh
pip install ultralytics==8.2.98安装好ultralytics后，后面的操纵需要更改源码，所以需要知道ultralytics安装的路径，可以使用下面的方式进行查询
sh
>>> import ultralytics
>>> ultralytics.__version__
'8.2.98'
>>> ultralytics.__path__
pathto\ultralytics修改
模型导出时的网络修改
修改pathto\ultralytics\nn\modules\head.py文件中的POSE类，在其forward函数中加如下代码
python
if self.export or torch.onnx.is_in_onnx_export():
results = self.forward_pose_export(x)
return tuple(results)同时在POSE类加上新如下函数
python
def forward_pose_export(self, x):
results = []
for i in range(self.nl):
dfl = self.cv2[i](x[i]).permute(0, 2, 3, 1).contiguous()
cls = self.cv3[i](x[i]).permute(0, 2, 3, 1).contiguous()
kpt = self.cv4[i](x[i]).permute(0, 2, 3, 1).contiguous()
results.append(torch.cat((cls, dfl, kpt), -1))
return results修改后的整体代码效果如下：
python
class Pose(Detect):
"""YOLOv8 Pose head for keypoints models."""
def __init__(self, nc=80, kpt_shape=(17, 3), ch=()):
"""Initialize YOLO network with default parameters and Convolutional Layers."""
super().__init__(nc, ch)
self.kpt_shape = kpt_shape # number of keypoints, number of dims (2 for x,y or 3 for x,y,visible)
self.nk = kpt_shape[0] * kpt_shape[1] # number of keypoints total
c4 = max(ch[0] // 4, self.nk)
self.cv4 = nn.ModuleList(nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, self.nk, 1)) for x in ch)
def forward(self, x):
"""Perform forward pass through YOLO model and return predictions."""
if self.export or torch.onnx.is_in_onnx_export():
results = self.forward_pose_export(x)
return tuple(results)
bs = x[0].shape[0] # batch size
kpt = torch.cat([self.cv4[i](x[i]).view(bs, self.nk, -1) for i in range(self.nl)], -1) # (bs, 17*3, h*w)
x = Detect.forward(self, x)
if self.training:
return x, kpt
pred_kpt = self.kpts_decode(bs, kpt)
return torch.cat([x, pred_kpt], 1) if self.export else (torch.cat([x[0], pred_kpt], 1), (x[1], kpt))
def forward_pose_export(self, x):
results = []
for i in range(self.nl):
dfl = self.cv2[i](x[i]).permute(0, 2, 3, 1).contiguous()
cls = self.cv3[i](x[i]).permute(0, 2, 3, 1).contiguous()
kpt = self.cv4[i](x[i]).permute(0, 2, 3, 1).contiguous()
results.append(torch.cat((cls, dfl, kpt), -1))
return results这种代码修改的导出方式并不影响训练的过程，仅在导出的时候会起到效果
更换激活函数并重头开始训练
YOLOv8默认使用的激活函数是SiLU,为了能更快的提升速度这里我换成了计算更高效的ReLU，并且更换激活函数后在转化成NCNN模型进行优化的时候会将卷积和ReLU融合之后,推理速度进一步得到了一些提升
更换激活函数后，需要对原有的Pytorch模型进行重新训练再导出ONNX
修改pathto\ultralytics\nn\modules\conv.py中的第39行左右的default_act = nn.SiLU() 修改为 default_act = nn.ReLU()
修改完成后使用训练脚本进行训练
python
from ultralytics import YOLO
model = YOLO('yolov8n-pose.yaml').load('yolov8n-pose.pt')
results = model.train(data='coco-pose.yaml', epochs=100, imgsz=640, workers=4, batch=64, project='Pose_runs', name='pose_n_relu')**注意： **
1、 .load('yolov8n-pose.pt')时加载预训练模型，虽然预训练的激活函数不一样，但是我测试了以下发现加上预训练模型前面的epoch的结果损失很
2、上面的数据使用的是coco2017数据集，训练环境为4090显卡，训练10个epoch接近一个小时
3、coco-pose.yaml、yolov8n-pose.yaml均是ultralytics默认配置文件
导出的ONNX名字修改
如果需要修改输出的名称则要去修改 pathto\ultralytics\engine\exporter.py 中的 export_onnx函数
ONNX导出
编写一个export.py的导出onnx的python代码
python
from ultralytics import YOLO
# load a pretrained model
model = YOLO('pathto\best.pt') # 训练得到的权重
# export onnx
model.export(format='onnx', opset=11, simplify=True, dynamic=False, imgsz=640)NCNN转化和优化
下面是onnx转化为NCNN的代码和对NCNN模型进行fp16的优化
sh
./onnx2ncnn yolov8n-pose.onnx yolov8n-pose-sim.param yolov8n-pose-sim.bin
./ncnnoptimize yolov8n-pose-sim.param yolov8n-pose-sim.bin yolov8n-pose-opt-fp16.param yolov8n-pose-opt-fp16.bin 1在使用ncnnoptimize对relu激活函数的onnx模型进行优化的时候，你会发现relu和卷积会进行算子融合
sh
(base) gx@RUKN0DC:/mnt/e/ubuntu_20.04/ncnn/build/install/bin$ ./ncnnoptimize yolov8pose-relu.param yolov8pose-relu.bin yolov8pose-relu-opt.param yolov8pose-relu-opt.bin 1
fuse_convolution_activation /model.0/conv/Conv /model.0/act/Relu
fuse_convolution_activation /model.1/conv/Conv /model.1/act/Relu
fuse_convolution_activation /model.2/cv1/conv/Conv /model.2/cv1/act/Relu
fuse_convolution_activation /model.2/m.0/cv1/conv/Conv /model.2/m.0/cv1/act/Relu
fuse_convolution_activation /model.2/m.0/cv2/conv/Conv /model.2/m.0/cv2/act/Relu
fuse_convolution_activation /model.2/cv2/conv/Conv /model.2/cv2/act/Relu
fuse_convolution_activation /model.3/conv/Conv /model.3/act/Relu
fuse_convolution_activation /model.4/cv1/conv/Conv /model.4/cv1/act/Relu
fuse_convolution_activation /model.4/m.0/cv1/conv/Conv /model.4/m.0/cv1/act/Relu
fuse_convolution_activation /model.4/m.0/cv2/conv/Conv /model.4/m.0/cv2/act/Relu
fuse_convolution_activation /model.4/m.1/cv1/conv/Conv /model.4/m.1/cv1/act/Relu
fuse_convolution_activation /model.4/m.1/cv2/conv/Conv /model.4/m.1/cv2/act/Relu
fuse_convolution_activation /model.4/cv2/conv/Conv /model.4/cv2/act/Relu
fuse_convolution_activation /model.5/conv/Conv /model.5/act/Relu
fuse_convolution_activation /model.6/cv1/conv/Conv /model.6/cv1/act/Relu
fuse_convolution_activation /model.6/m.0/cv1/conv/Conv /model.6/m.0/cv1/act/Relu
fuse_convolution_activation /model.6/m.0/cv2/conv/Conv /model.6/m.0/cv2/act/Relu
fuse_convolution_activation /model.6/m.1/cv1/conv/Conv /model.6/m.1/cv1/act/Relu
fuse_convolution_activation /model.6/m.1/cv2/conv/Conv /model.6/m.1/cv2/act/Relu
fuse_convolution_activation /model.6/cv2/conv/Conv /model.6/cv2/act/Relu
fuse_convolution_activation /model.7/conv/Conv /model.7/act/Relu
fuse_convolution_activation /model.8/cv1/conv/Conv /model.8/cv1/act/Relu
fuse_convolution_activation /model.8/m.0/cv1/conv/Conv /model.8/m.0/cv1/act/Relu
fuse_convolution_activation /model.8/m.0/cv2/conv/Conv /model.8/m.0/cv2/act/Relu
fuse_convolution_activation /model.8/cv2/conv/Conv /model.8/cv2/act/Relu
fuse_convolution_activation /model.9/cv1/conv/Conv /model.9/cv1/act/Relu
fuse_convolution_activation /model.9/cv2/conv/Conv /model.9/cv2/act/Relu
fuse_convolution_activation /model.12/cv1/conv/Conv /model.12/cv1/act/Relu
fuse_convolution_activation /model.12/m.0/cv1/conv/Conv /model.12/m.0/cv1/act/Relu
fuse_convolution_activation /model.12/m.0/cv2/conv/Conv /model.12/m.0/cv2/act/Relu
fuse_convolution_activation /model.12/cv2/conv/Conv /model.12/cv2/act/Relu
fuse_convolution_activation /model.15/cv1/conv/Conv /model.15/cv1/act/Relu
fuse_convolution_activation /model.15/m.0/cv1/conv/Conv /model.15/m.0/cv1/act/Relu
fuse_convolution_activation /model.15/m.0/cv2/conv/Conv /model.15/m.0/cv2/act/Relu
fuse_convolution_activation /model.15/cv2/conv/Conv /model.15/cv2/act/Relu
fuse_convolution_activation /model.16/conv/Conv /model.16/act/Relu
fuse_convolution_activation /model.18/cv1/conv/Conv /model.18/cv1/act/Relu
fuse_convolution_activation /model.18/m.0/cv1/conv/Conv /model.18/m.0/cv1/act/Relu
fuse_convolution_activation /model.18/m.0/cv2/conv/Conv /model.18/m.0/cv2/act/Relu
fuse_convolution_activation /model.18/cv2/conv/Conv /model.18/cv2/act/Relu
fuse_convolution_activation /model.19/conv/Conv /model.19/act/Relu
fuse_convolution_activation /model.21/cv1/conv/Conv /model.21/cv1/act/Relu
fuse_convolution_activation /model.21/m.0/cv1/conv/Conv /model.21/m.0/cv1/act/Relu
fuse_convolution_activation /model.21/m.0/cv2/conv/Conv /model.21/m.0/cv2/act/Relu
fuse_convolution_activation /model.21/cv2/conv/Conv /model.21/cv2/act/Relu
fuse_convolution_activation /model.22/cv2.0/cv2.0.0/conv/Conv /model.22/cv2.0/cv2.0.0/act/Relu
fuse_convolution_activation /model.22/cv2.0/cv2.0.1/conv/Conv /model.22/cv2.0/cv2.0.1/act/Relu
fuse_convolution_activation /model.22/cv3.0/cv3.0.0/conv/Conv /model.22/cv3.0/cv3.0.0/act/Relu
fuse_convolution_activation /model.22/cv3.0/cv3.0.1/conv/Conv /model.22/cv3.0/cv3.0.1/act/Relu
fuse_convolution_activation /model.22/cv4.0/cv4.0.0/conv/Conv /model.22/cv4.0/cv4.0.0/act/Relu
fuse_convolution_activation /model.22/cv4.0/cv4.0.1/conv/Conv /model.22/cv4.0/cv4.0.1/act/Relu
fuse_convolution_activation /model.22/cv2.1/cv2.1.0/conv/Conv /model.22/cv2.1/cv2.1.0/act/Relu
fuse_convolution_activation /model.22/cv2.1/cv2.1.1/conv/Conv /model.22/cv2.1/cv2.1.1/act/Relu
fuse_convolution_activation /model.22/cv3.1/cv3.1.0/conv/Conv /model.22/cv3.1/cv3.1.0/act/Relu
fuse_convolution_activation /model.22/cv3.1/cv3.1.1/conv/Conv /model.22/cv3.1/cv3.1.1/act/Relu
fuse_convolution_activation /model.22/cv4.1/cv4.1.0/conv/Conv /model.22/cv4.1/cv4.1.0/act/Relu
fuse_convolution_activation /model.22/cv4.1/cv4.1.1/conv/Conv /model.22/cv4.1/cv4.1.1/act/Relu
fuse_convolution_activation /model.22/cv2.2/cv2.2.0/conv/Conv /model.22/cv2.2/cv2.2.0/act/Relu
fuse_convolution_activation /model.22/cv2.2/cv2.2.1/conv/Conv /model.22/cv2.2/cv2.2.1/act/Relu
fuse_convolution_activation /model.22/cv3.2/cv3.2.0/conv/Conv /model.22/cv3.2/cv3.2.0/act/Relu
fuse_convolution_activation /model.22/cv3.2/cv3.2.1/conv/Conv /model.22/cv3.2/cv3.2.1/act/Relu
fuse_convolution_activation /model.22/cv4.2/cv4.2.0/conv/Conv /model.22/cv4.2/cv4.2.0/act/Relu
fuse_convolution_activation /model.22/cv4.2/cv4.2.1/conv/Conv /model.22/cv4.2/cv4.2.1/act/Relu
Input layer images without shape info, shape_inference skipped
Input layer images without shape info, estimate_memory_footprint skipped下面是ONNX、NCNN模型以及NCNN优化后的三个模型结构简单对比
下面是不同的激活函数最后NCNN优化后的网络结构简单对比
安卓代码的修改
参考这两个代码进行修改
https://github.com/eecn/ncnn-android-yolov8-pose
https://github.com/Rachel-liuqr/yolov8s-pose-ncnn
有以下几个修改的地方：
- 将sigmoid函数修改为了使用快速指数fast_exp的sigmoid
- 将 cv::dnn::NMSBoxes 修改了使用纯C++代码的实现
具体的代码过程，有兴趣的可以去查看
本人水平高，代码应该还有更大的优化空间！！
参考资料
https://github.com/eecn/ncnn-android-yolov8-pose
https://blog.csdn.net/Rachel321/article/details/130381788