mirror of
https://github.com/PaddlePaddle/FastDeploy.git
synced 2025-10-31 11:56:44 +08:00
79d0ee07c9
* renew develop_a_new_model * renew doc * update readme file * fix review problem * fix review problem --------- Co-authored-by: WJJ1995 <wjjisloser@163.com>
387 lines
16 KiB
Markdown
387 lines
16 KiB
Markdown
|
||
[English](../../en/faq/develop_a_new_model.md) | 中文
|
||
|
||
# FastDeploy集成新模型流程
|
||
|
||
在FastDeploy里面新增一个模型，包括增加C++/Python的部署支持。 本文以YOLOv7Face模型为例，介绍使用FastDeploy做外部[模型集成](#modelsupport)，具体包括如下3步。
|
||
|
||
| 步骤 | 说明 | 创建或修改的文件 |
|
||
|:------:|:-------------------------------------:|:---------------------------------------------:|
|
||
| [1](#step2) | 在fastdeploy/vision相应任务模块增加模型实现 | yolov7face.h、yolov7face.cc、preprocessor.h、preprocess.cc、postprocessor.h、postprocessor.cc、vision.h |
|
||
| [2](#step4) | 通过pybind完成Python接口绑定 | yolov7face_pybind.cc |
|
||
| [3](#step5) | 实现Python相应调用接口 | yolov7face.py、\_\_init\_\_.py |
|
||
|
||
在完成上述3步之后，一个外部模型就集成好了。
|
||
<br />
|
||
如果您想为FastDeploy贡献代码，还需要为新增模型添加测试代码、说明文档和代码注释，可在[测试](#test)中查看。
|
||
## 模型集成 <span id="modelsupport"></span>
|
||
|
||
## 1、模型准备 <span id="step1"></span>
|
||
|
||
在集成外部模型之前，先要将训练好的模型（.pt，.pdparams 等）转换成FastDeploy支持部署的模型格式（.onnx，.pdmodel）。多数开源仓库会提供模型转换脚本，可以直接利用脚本做模型的转换。例如yolov7face官方库提供的[export.py](https://github.com/derronqi/yolov7-face/blob/main/models/export.py)文件， 若官方库未提供转换导出文件，则需要手动编写转换脚本，如torchvision没有提供转换脚本，因此手动编写转换脚本，下文中将 `torchvison.models.resnet50` 转换为 `resnet50.onnx`，参考代码如下：
|
||
|
||
```python
|
||
import torch
|
||
import torchvision.models as models
|
||
model = models.resnet50(pretrained=True)
|
||
batch_size = 1 #批处理大小
|
||
input_shape = (3, 224, 224) #输入数据,改成自己的输入shape
|
||
model.eval()
|
||
x = torch.randn(batch_size, *input_shape) # 生成张量
|
||
export_onnx_file = "resnet50.onnx" # 目的ONNX文件名
|
||
torch.onnx.export(model,
|
||
x,
|
||
export_onnx_file,
|
||
opset_version=12,
|
||
input_names=["input"], # 输入名
|
||
output_names=["output"], # 输出名
|
||
dynamic_axes={"input":{0:"batch_size"}, # 批处理变量
|
||
"output":{0:"batch_size"}})
|
||
```
|
||
执行上述脚本将会得到 `resnet50.onnx` 文件。
|
||
|
||
## 2、CPP代码实现 <span id="step2"></span>
|
||
### 2.1、前处理类实现
|
||
* 创建`preprocessor.h`文件
|
||
* 创建位置
|
||
* FastDeploy/fastdeploy/vision/facedet/contrib/yolov7face/preprocess.h (FastDeploy/C++代码存放位置/视觉模型/任务名称/外部模型/模型名/precessor.h)
|
||
* 创建内容
|
||
* 首先在preprocess.h中创建 Yolov7FacePreprocess 类,之后声明`Run`、`preprocess`、`LetterBox`和`构造函数`，以及必要的变量及其`set`和`get`方法，具体的代码细节请参考[preprocess.h](https://github.com/PaddlePaddle/FastDeploy/tree/develop/fastdeploy/vision/facedet/contrib/yolov7face/preprocessor.h)。
|
||
|
||
```C++
|
||
class FASTDEPLOY_DECL Yolov7FacePreprocessor {
|
||
public:
|
||
Yolov7FacePreprocessor(...);
|
||
bool Run(...);
|
||
protected:
|
||
bool Preprocess(...);
|
||
void LetterBox(...);
|
||
};
|
||
```
|
||
|
||
* 创建`preprocessor.cc`文件
|
||
* 创建位置
|
||
* FastDeploy/fastdeploy/vision/facedet/contrib/yolov7face/preprocessor.cc (FastDeploy/C++代码存放位置/视觉模型/任务名称/外部模型/模型名/preprocessor.cc)
|
||
* 创建内容
|
||
* 在`preprocessor.cc`中实现`preprocessor.h`中声明函数的具体逻辑，其中`Preprocess`需要参考源官方库的前后处理逻辑复现，preprocessor每个函数具体逻辑如下，具体的代码请参考[preprocessor.cc](https://github.com/PaddlePaddle/FastDeploy/tree/develop/fastdeploy/vision/facedet/contrib/yolov7face/preprocessor.cc)。
|
||
|
||
```C++
|
||
Yolov7FacePreprocessor::Yolov7FacePreprocessor(...) {
|
||
// 构造函数逻辑
|
||
// 全局变量赋值
|
||
}
|
||
bool Yolov7FacePreprocessor::Run() {
|
||
// 执行前处理
|
||
// 根据传入图片数量对每张图片进行处理，通过循环的方式将每张图片传入Preprocess函数进行预处理,
|
||
// 即Preprocess为处理单元，Run方法为每张图片调用处理单元处理
|
||
return true;
|
||
}
|
||
bool Yolov7FacePreprocessor::Preprocess(FDMat* mat, FDTensor* output,
|
||
std::map<std::string, std::array<float, 2>>* im_info) {
|
||
// 前处理逻辑
|
||
// 1. LetterBox 2. convert and permute 3. 处理结果存入 FDTensor类中
|
||
return true;
|
||
}
|
||
void Yolov7FacePreprocessor::LetterBox(FDMat* mat) {
|
||
//LetterBox
|
||
return true;
|
||
}
|
||
```
|
||
|
||
### 2.2、后处理类实现
|
||
* 创建`postprocessor.h`文件
|
||
* 创建位置
|
||
* FastDeploy/fastdeploy/vision/facedet/contrib/yolov7face/postprocessor.h (FastDeploy/C++代码存放位置/视觉模型/任务名称/外部模型/模型名/postprocessor.h)
|
||
* 创建内容
|
||
* 首先在postprocess.h中创建 Yolov7FacePostprocess 类,之后声明`Run`和`构造函数`，以及必要的变量及其`set`和`get`方法，具体的代码细节请参考[postprocessor.h](https://github.com/PaddlePaddle/FastDeploy/tree/develop/fastdeploy/vision/facedet/contrib/yolov7face/postprocessor.h)。
|
||
|
||
```C++
|
||
class FASTDEPLOY_DECL Yolov7FacePostprocessor {
|
||
public:
|
||
Yolov7FacePostprocessor(...);
|
||
bool Run(...);
|
||
};
|
||
```
|
||
|
||
* 创建`postprocessor.cc`文件
|
||
* 创建位置
|
||
* FastDeploy/fastdeploy/vision/facedet/contrib/yolov7face/postprocessor.cc (FastDeploy/C++代码存放位置/视觉模型/任务名称/外部模型/模型名/postprocessor.cc)
|
||
* 创建内容
|
||
* 在`postprocessor.cc`中实现`postprocessor.h`中声明函数的具体逻辑，其中`Postprocess`需要参考源官方库的前后处理逻辑复现，postprocessor每个函数具体逻辑如下，具体的代码请参考[postprocessor.cc](https://github.com/PaddlePaddle/FastDeploy/tree/develop/fastdeploy/vision/facedet/contrib/yolov7face/postprocessor.cc)。
|
||
|
||
```C++
|
||
Yolov7FacePostprocessor::Yolov7FacePostprocessor(...) {
|
||
// 构造函数逻辑
|
||
// 全局变量赋值
|
||
}
|
||
bool Yolov7FacePostprocessor::Run() {
|
||
// 后处理逻辑
|
||
// 1. Padding 2. Choose box by conf_threshold 3. NMS 4. 结果存入 FaceDetectionResult类
|
||
return true;
|
||
}
|
||
|
||
```
|
||
### 2.3、YOLOv7Face实现
|
||
* 创建`yolov7face.h`文件
|
||
* 创建位置
|
||
* FastDeploy/fastdeploy/vision/facedet/contrib/yolov7face/yolov7face.h (FastDeploy/C++代码存放位置/视觉模型/任务名称/外部模型/模型名/模型名.h)
|
||
* 创建内容
|
||
* 首先在yolov7face.h中创建 YOLOv7Face 类并继承FastDeployModel父类，之后声明`Predict`、`BatchPredict`、`Initialize`和`构造函数`，以及必要的变量及其`get`方法，具体的代码细节请参考[yolov7face.h](https://github.com/PaddlePaddle/FastDeploy/tree/develop/fastdeploy/vision/facedet/contrib/yolov7face/yolov7face.h)。
|
||
|
||
```C++
|
||
class FASTDEPLOY_DECL YOLOv7Face : public FastDeployModel {
|
||
public:
|
||
YOLOv7Face(...);
|
||
virtual bool Predict(...);
|
||
virtual bool BatchPredict(...);
|
||
protected:
|
||
bool Initialize();
|
||
Yolov7FacePreprocessor preprocessor_;
|
||
Yolov7FacePostprocessor postprocessor_;
|
||
};
|
||
```
|
||
|
||
* 创建`yolov7face.cc`文件
|
||
* 创建位置
|
||
* FastDeploy/fastdeploy/vision/facedet/contrib/yolov7face/yolov7face.cc (FastDeploy/C++代码存放位置/视觉模型/任务名称/外部模型/模型名/模型名.cc)
|
||
* 创建内容
|
||
* 在`yolov7face.cc`中实现`yolov7face.h`中声明函数的具体逻辑，YOLOv7Face每个函数具体逻辑如下，具体的代码请参考[yolov7face.cc](https://github.com/PaddlePaddle/FastDeploy/tree/develop/fastdeploy/vision/facedet/contrib/yolov7face/yolov7face.cc)。
|
||
|
||
```C++
|
||
YOLOv7Face::YOLOv7Face(...) {
|
||
// 构造函数逻辑
|
||
// 1. 指定 Backend 2. 设置RuntimeOption 3. 调用Initialize()函数
|
||
}
|
||
bool YOLOv7Face::Initialize() {
|
||
// 初始化逻辑
|
||
// 1. 全局变量赋值 2. 调用InitRuntime()函数
|
||
return true;
|
||
}
|
||
bool YOLOv7Face::Predict(const cv::Mat& im, FaceDetectionResult* result) {
|
||
std::vector<FaceDetectionResult> results;
|
||
if (!BatchPredict({im}, &results)) {
|
||
return false;
|
||
}
|
||
*result = std::move(results[0]);
|
||
return true;
|
||
}
|
||
// Predict是对单张图片进行预测，通过将含有一张图片的数组送入BatchPredict实现
|
||
bool YOLOv7Face::BatchPredict(const std::vector<cv::Mat>& images, std::vector<FaceDetectionResult>* result) {
|
||
Preprocess(...)
|
||
Infer(...)
|
||
Postprocess(...)
|
||
return true;
|
||
}
|
||
// BatchPredict为对批量图片进行预测，接收一个含有若干张图片的动态数组vector
|
||
```
|
||
<span id="step3"></span>
|
||
* 在`vision.h`文件中加入新增模型文件
|
||
* 修改位置
|
||
* FastDeploy/fastdeploy/vision.h
|
||
* 修改内容
|
||
|
||
```C++
|
||
#ifdef ENABLE_VISION
|
||
#include "fastdeploy/vision/facedet/contrib/yolov7face.h"
|
||
#endif
|
||
```
|
||
|
||
## 3、Python接口封装
|
||
|
||
### 3.1、Pybind部分 <span id="step4"></span>
|
||
|
||
* 创建Pybind文件
|
||
* 创建位置
|
||
* FastDeploy/fastdeploy/vision/facedet/contrib/yolov7face/yolov7face_pybind.cc (FastDeploy/C++代码存放位置/视觉模型/任务名称/外部模型/模型名/模型名_pybind.cc)
|
||
* 创建内容
|
||
* 利用Pybind将C++中的函数变量绑定到Python中，具体代码请参考[yolov7face_pybind.cc](https://github.com/PaddlePaddle/FastDeploy/tree/develop/fastdeploy/vision/facedet/contrib/yolov7face/yolov7face_pybind.cc)。
|
||
```C++
|
||
void BindYOLOv7Face(pybind11::module& m) {
|
||
pybind11::class_<vision::facedet::YOLOv7Face, FastDeployModel>(
|
||
m, "YOLOv7Face")
|
||
.def(pybind11::init<std::string, std::string, RuntimeOption, ModelFormat>())
|
||
.def("predict", ...)
|
||
.def("batch_predict", ...)
|
||
.def_property_readonly("preprocessor", ...)
|
||
.def_property_readonly("postprocessor", ...);
|
||
pybind11::class_<vision::facedet::Yolov7FacePreprocessor>(
|
||
m, "Yolov7FacePreprocessor")
|
||
.def(pybind11::init<>())
|
||
.def("run", ...)
|
||
.def_property("size", ...)
|
||
.def_property("padding_color_value", ...)
|
||
.def_property("is_scale_up", ...);
|
||
pybind11::class_<vision::facedet::Yolov7FacePostprocessor>(
|
||
m, "Yolov7FacePostprocessor")
|
||
.def(pybind11::init<>())
|
||
.def("run", ...)
|
||
.def_property("conf_threshold", ...)
|
||
.def_property("nms_threshold", ...);
|
||
}
|
||
```
|
||
|
||
* 调用Pybind函数
|
||
* 修改位置
|
||
* FastDeploy/fastdeploy/vision/facedet/facedet_pybind.cc (FastDeploy/C++代码存放位置/视觉模型/任务名称/任务名称}_pybind.cc)
|
||
* 修改内容
|
||
```C++
|
||
void BindYOLOv7Face(pybind11::module& m);
|
||
void BindFaceDet(pybind11::module& m) {
|
||
auto facedet_module =
|
||
m.def_submodule("facedet", "Face detection models.");
|
||
BindYOLOv7Face(facedet_module);
|
||
}
|
||
```
|
||
|
||
### 3.2、python部分 <span id="step5"></span>
|
||
* 创建`yolov7face.py`文件
|
||
* 创建位置
|
||
* FastDeploy/python/fastdeploy/vision/facedet/contrib/yolov7face.py (FastDeploy/Python代码存放位置/fastdeploy/视觉模型/任务名称/外部模型/模型名.py)
|
||
* 创建内容
|
||
* 创建YOLOv7Face类继承自FastDeployModel、preprocess以及postprocess类，实现 `\_\_init\_\_`、Pybind绑定的函数（如`predict()`）、以及`对Pybind绑定的全局变量进行赋值和获取的函数`，具体代码请参考[yolov7face.py](https://github.com/PaddlePaddle/FastDeploy/tree/develop/python/fastdeploy/vision/facedet/contrib/yolov7face.py)。
|
||
|
||
```python
|
||
class YOLOv7Face(FastDeployModel):
|
||
def __init__(self, ...):
|
||
self._model = C.vision.facedet.YOLOv7Face(...)
|
||
def predict(self, input_image):
|
||
return self._model.predict(input_image)
|
||
def batch_predict(self, images):
|
||
return self._model.batch_predict(images)
|
||
@property
|
||
def preprocessor(self):
|
||
return self._model.preprocessor
|
||
@property
|
||
def postprocessor(self):
|
||
return self._model.postprocessor
|
||
|
||
class Yolov7FacePreprocessor():
|
||
def __init__(self, ...):
|
||
self._model = C.vision.facedet.Yolov7FacePreprocessor(...)
|
||
def run(self, input_ims):
|
||
return self._preprocessor.run(input_ims)
|
||
@property
|
||
def size(self):
|
||
return self._preprocessor.size
|
||
@property
|
||
def padding_color_value(self):
|
||
return self._preprocessor.padding_color_value
|
||
...
|
||
|
||
class Yolov7FacePreprocessor():
|
||
def __init__(self, ...):
|
||
self._model = C.vision.facedet.Yolov7FacePostprocessor(...)
|
||
def run(self, ...):
|
||
return self._postprocessor.run(...)
|
||
@property
|
||
def conf_threshold(self):
|
||
return self._postprocessor.conf_threshold
|
||
@property
|
||
def nms_threshold(self):
|
||
return self._postprocessor.nms_threshold
|
||
...
|
||
```
|
||
<span id="step6"></span>
|
||
* 导入YOLOv7Face、Yolov7FacePreprocessor、Yolov7facePostprocessor类
|
||
* 修改位置
|
||
* FastDeploy/python/fastdeploy/vision/facedet/\_\_init\_\_.py (FastDeploy/Python代码存放位置/fastdeploy/视觉模型/任务名称/\_\_init\_\_.py)
|
||
* 修改内容
|
||
|
||
```Python
|
||
from .contrib.yolov7face import *
|
||
```
|
||
|
||
## 4、测试 <span id="test"></span>
|
||
### 编译
|
||
* C++
|
||
* 位置：FastDeploy/
|
||
|
||
```
|
||
mkdir build & cd build
|
||
cmake .. -DENABLE_ORT_BACKEND=ON -DENABLE_VISION=ON -DCMAKE_INSTALL_PREFIX=${PWD/fastdeploy-0.0.3
|
||
-DENABLE_PADDLE_BACKEND=ON -DENABLE_TRT_BACKEND=ON -DWITH_GPU=ON -DTRT_DIRECTORY=/PATH/TO/TensorRT/
|
||
make -j8
|
||
make install
|
||
```
|
||
|
||
编译会得到 build/fastdeploy-0.0.3/。
|
||
|
||
* Python
|
||
* 位置：FastDeploy/python/
|
||
|
||
```
|
||
export TRT_DIRECTORY=/PATH/TO/TensorRT/ # 如果用TensorRT 需要填写TensorRT所在位置，并开启 ENABLE_TRT_BACKEND
|
||
export ENABLE_TRT_BACKEND=ON
|
||
export WITH_GPU=ON
|
||
export ENABLE_PADDLE_BACKEND=ON
|
||
export ENABLE_OPENVINO_BACKEND=ON
|
||
export ENABLE_VISION=ON
|
||
export ENABLE_ORT_BACKEND=ON
|
||
python setup.py build
|
||
python setup.py bdist_wheel
|
||
cd dist
|
||
pip install fastdeploy_gpu_python-版本号-cpxx-cpxxm-系统架构.whl
|
||
```
|
||
|
||
## 5、示例代码开发
|
||
* 创建位置: FastDeploy/examples/vision/facedet/yolov7face/ (FastDeploy/示例目录/视觉模型/任务名称/模型名/)
|
||
* 创建目录结构
|
||
|
||
```
|
||
.
|
||
├── cpp
|
||
│ ├── CMakeLists.txt
|
||
│ ├── infer.cc // C++ 版本测试代码
|
||
│ └── README.md // C++版本使用文档
|
||
├── python
|
||
│ ├── infer.py // Python 版本测试代码
|
||
│ └── README.md // Python版本使用文档
|
||
└── README.md // ResNet 模型集成说明文档
|
||
```
|
||
|
||
* C++
|
||
* 编写CmakeLists文件、C++ 代码以及 README.md 内容请参考[cpp/](https://github.com/PaddlePaddle/FastDeploy/tree/develop/examples/vision/facedet/yolov7face/cpp)。
|
||
* 编译 infer.cc
|
||
* 位置：FastDeploy/examples/vision/facedet/yolov7face/cpp/
|
||
|
||
```
|
||
mkdir build & cd build
|
||
cmake .. -DFASTDEPLOY_INSTALL_DIR=/PATH/TO/FastDeploy/build/fastdeploy-0.0.3/
|
||
make
|
||
```
|
||
|
||
* Python
|
||
* Python 代码以及 README.md 内容请参考[python/](https://github.com/PaddlePaddle/FastDeploy/tree/develop/examples/vision/facedet/yolov7face/python)。
|
||
|
||
### 为代码添加注释
|
||
为了方便用户理解代码，我们需要为新增代码添加注释，添加注释方法可参考如下示例。
|
||
- C++ 代码
|
||
您需要在resnet.h文件中为函数和变量增加注释，有如下三种注释方式，具体可参考[yolov7face.h](https://github.com/PaddlePaddle/FastDeploy/tree/develop/fastdeploy/vision/facedet/contrib/yolov7face/yolov7face.h)。
|
||
|
||
```C++
|
||
/** \brief Predict for the input "im", the result will be saved in "result".
|
||
*
|
||
* \param[in] im Input image for inference.
|
||
* \param[in] result Saving the inference result.
|
||
*/
|
||
virtual bool Predict(const cv::Mat& im, FaceDetectionResult* result);
|
||
/// Tuple of (width, height)
|
||
std::vector<int> size;
|
||
/*! @brief Initialize for YOLOv7Face model, assign values to the global variables and call InitRuntime()
|
||
*/
|
||
bool Initialize();
|
||
```
|
||
- Python 代码
|
||
你需要为yolov7face.py文件中的函数和变量增加适当的注释，示例如下，具体可参考[yolov7face.py](https://github.com/PaddlePaddle/FastDeploy/tree/develop/python/fastdeploy/vision/facedet/contrib/yolov7face.py)。
|
||
|
||
```python
|
||
def predict(self, input_image):
|
||
"""Detect the location and key points of human faces from an input image
|
||
:param input_image: (numpy.ndarray)The input image data, 3-D array with layout HWC, BGR format
|
||
:return: FaceDetectionResult
|
||
"""
|
||
return self._model.predict(input_image)
|
||
```
|
||
|
||
对于集成模型过程中的其他文件，您也可以对实现的细节添加适当的注释说明。
|