[Model] Add tinypose single && pipeline model (#177)

* Add tinypose model * Add PPTinypose python API * Fix picodet preprocess bug && Add Tinypose examples * Update tinypose example code * Update ppseg preprocess if condition * Update ppseg backend support type * Update permute.h * Update README.md * Update code with comments * Move files dir * Delete premute.cc * Add single model pptinypose * Delete pptinypose old code in ppdet * Code format * Add ppdet + pptinypose pipeline model * Fix bug for posedetpipeline * Change Frontend to ModelFormat * Change Frontend to ModelFormat in __init__.py * Add python posedetpipeline/ * Update pptinypose example dir name * Update README.md * Update README.md * Update README.md * Update README.md * Create keypointdetection_result.md * Create README.md * Create README.md * Create README.md * Update README.md * Update README.md * Create README.md * Fix det_keypoint_unite_infer.py bug * Create README.md * Update PP-Tinypose by comment * Update by comment * Add pipeline directory * Add pptinypose dir * Update pptinypose to align accuracy * Addd warpAffine processor * Update GetCpuMat to GetOpenCVMat * Add comment for pptinypose && pipline * Update docs/main_page.md * Add README.md for pptinypose * Add README for det_keypoint_unite * Remove ENABLE_PIPELINE option * Remove ENABLE_PIPELINE option * Change pptinypose default backend * PP-TinyPose Pipeline support multi PP-Detection models * Update pp-tinypose comment * Update by comments * Add single test example Co-authored-by: Jason <jiangjiajun@baidu.com>
2025-10-26 01:50:33 +08:00 · 2022-10-21 09:28:23 +08:00
parent 49ab773d22
commit b565c15bf7
62 changed files with 2583 additions and 20 deletions
--- a/examples/vision/keypointdetection/tiny_pose/README.md
+++ b/examples/vision/keypointdetection/tiny_pose/README.md
@@ -0,0 +1,37 @@
+# PP-TinyPose 模型部署
+
+## 模型版本说明
+
+- [PaddleDetection release/2.5](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.5)
+
+目前FastDeploy支持如下模型的部署 
+
+- [PP-TinyPose系列模型](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.5/configs/keypoint/tiny_pose/README.md)
+
+## 准备PP-TinyPose部署模型
+
+PP-TinyPose模型导出，请参考其文档说明[模型导出](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.5/deploy/EXPORT_MODEL.md)  
+
+**注意**:PP-TinyPose导出的模型包含`model.pdmodel`、`model.pdiparams`和`infer_cfg.yml`三个文件，FastDeploy会从yaml文件中获取模型在推理时需要的预处理信息。
+
+
+## 下载预训练模型
+
+为了方便开发者的测试，下面提供了PP-TinyPose导出的部分模型，开发者可直接下载使用。
+
+| 模型                                                               | 参数文件大小 |输入Shape |  AP(业务数据集) | AP(COCO Val) | FLOPS | 单人推理耗时 (FP32) | 单人推理耗时（FP16) |
+|:---------------------------------------------------------------- |:----- |:----- | :----- | :----- | :----- | :----- | :----- |
+| [PP-TinyPose-128x96](https://bj.bcebos.com/paddlehub/fastdeploy/PP_TinyPose_128x96_infer.tgz) | 5.3MB | 128x96 | 84.3% | 58.4% | 81.56 M | 4.57ms | 3.27ms |
+| [PP-TinyPose-256x192](https://bj.bcebos.com/paddlehub/fastdeploy/PP_TinyPose_256x192_infer.tgz)  | 5.3M  | 256x96 | 91.0% | 68.3% | 326.24M | 14.07ms | 8.33ms |
+
+**说明**
+- 关键点检测模型使用`COCO train2017`和`AI Challenger trainset`作为训练集。使用`COCO person keypoints val2017`作为测试集。
+- 关键点检测模型的精度指标所依赖的检测框为ground truth标注得到。
+- 推理速度测试环境为 Qualcomm Snapdragon 865，采用arm8下4线程推理得到。
+
+更多信息请参考：[PP-TinyPose 官方文档](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.5/configs/keypoint/tiny_pose/README.md)
+
+## 详细部署文档
+
+- [Python部署](python)
+- [C++部署](cpp)
--- a/examples/vision/keypointdetection/tiny_pose/cpp/CMakeLists.txt
+++ b/examples/vision/keypointdetection/tiny_pose/cpp/CMakeLists.txt
@@ -0,0 +1,13 @@
+PROJECT(infer_demo C CXX)
+CMAKE_MINIMUM_REQUIRED (VERSION 3.12)
+
+# 指定下载解压后的fastdeploy库路径
+option(FASTDEPLOY_INSTALL_DIR "Path of downloaded fastdeploy sdk.")
+
+include(${FASTDEPLOY_INSTALL_DIR}/FastDeploy.cmake)
+
+# 添加FastDeploy依赖头文件
+include_directories(${FASTDEPLOY_INCS})
+
+add_executable(infer_tinypose_demo ${PROJECT_SOURCE_DIR}/pptinypose_infer.cc)
+target_link_libraries(infer_tinypose_demo ${FASTDEPLOY_LIBS})
--- a/examples/vision/keypointdetection/tiny_pose/cpp/README.md
+++ b/examples/vision/keypointdetection/tiny_pose/cpp/README.md
@@ -0,0 +1,88 @@
+# PP-TinyPose C++部署示例
+
+本目录下提供`pptinypose_infer.cc`快速完成PP-TinyPose在CPU/GPU，以及GPU上通过TensorRT加速部署的`单图单人关键点检测`示例
+>> **注意**: PP-Tinypose单模型目前只支持单图单人关键点检测，因此输入的图片应只包含一个人或者进行过裁剪的图像。多人关键点检测请参考[PP-TinyPose Pipeline](../../det_keypoint_unite/cpp/README.md)
+
+在部署前，需确认以下两个步骤
+
+- 1. 软硬件环境满足要求，参考[FastDeploy环境要求](../../../../../docs/cn/build_and_install/download_prebuilt_libraries.md)  
+- 2. 根据开发环境，下载预编译部署库和samples代码，参考[FastDeploy预编译库](../../../../../docs/cn/build_and_install/download_prebuilt_libraries.md)
+
+
+以Linux上推理为例，在本目录执行如下命令即可完成编译测试
+
+```bash
+wget https://bj.bcebos.com/fastdeploy/release/cpp/fastdeploy-linux-x64-gpu-0.3.0.tgz
+tar xvf fastdeploy-linux-x64-gpu-0.3.0.tgz
+cd fastdeploy-linux-x64-gpu-0.3.0/examples/vision/keypointdetection/tiny_pose/cpp/
+mkdir build
+cd build
+cmake .. -DFASTDEPLOY_INSTALL_DIR=${PWD}/../../../../../../../fastdeploy-linux-x64-gpu-0.3.0
+make -j
+
+# 下载PP-TinyPose模型文件和测试图片
+wget https://bj.bcebos.com/paddlehub/fastdeploy/PP_TinyPose_256x192_infer.tgz
+tar -xvf PP_TinyPose_256x192_infer.tgz
+wget https://bj.bcebos.com/paddlehub/fastdeploy/hrnet_demo.jpg
+
+
+# CPU推理
+./infer_tinypose_demo PP_TinyPose_256x192_infer hrnet_demo.jpg 0
+# GPU推理
+./infer_tinypose_demo PP_TinyPose_256x192_infer hrnet_demo.jpg 1
+# GPU上TensorRT推理
+./infer_tinypose_demo PP_TinyPose_256x192_infer hrnet_demo.jpg 2
+```
+
+运行完成可视化结果如下图所示
+<div  align="center">  
+<img src="https://user-images.githubusercontent.com/16222477/196386764-dd51ad56-c410-4c54-9580-643f282f5a83.jpeg", width=359px, height=423px />
+</div>
+
+以上命令只适用于Linux或MacOS, Windows下SDK的使用方式请参考:  
+- [如何在Windows中使用FastDeploy C++ SDK](../../../../../docs/cn/faq/use_sdk_on_windows.md)
+
+## PP-TinyPose C++接口
+
+### PP-TinyPose类
+
+```c++
+fastdeploy::vision::keypointdetection::PPTinyPose(
+        const string& model_file,
+        const string& params_file = "",
+        const string& config_file,
+        const RuntimeOption& runtime_option = RuntimeOption(),
+        const ModelFormat& model_format = ModelFormat::PADDLE)
+```
+
+PPTinyPose模型加载和初始化，其中model_file为导出的Paddle模型格式。
+
+**参数**
+
+> * **model_file**(str): 模型文件路径
+> * **params_file**(str): 参数文件路径
+> * **config_file**(str): 推理部署配置文件
+> * **runtime_option**(RuntimeOption): 后端推理配置，默认为None，即采用默认配置
+> * **model_format**(ModelFormat): 模型格式，默认为Paddle格式
+
+#### Predict函数
+
+> ```c++
+> PPTinyPose::Predict(cv::Mat* im, KeyPointDetectionResult* result)
+> ```
+>
+> 模型预测接口，输入图像直接输出关键点检测结果。
+>
+> **参数**
+>
+> > * **im**: 输入图像，注意需为HWC，BGR格式
+> > * **result**: 关键点检测结果，包括关键点的坐标以及关键点对应的概率值, KeyPointDetectionResult说明参考[视觉模型预测结果](../../../../../docs/api/vision_results/)
+
+### 类成员属性
+#### 后处理参数
+> > * **use_dark**(bool): 是否使用DARK进行后处理[参考论文](https://arxiv.org/abs/1910.06278)
+
+- [模型介绍](../../)
+- [Python部署](../python)
+- [视觉模型预测结果](../../../../../docs/api/vision_results/)
+- [如何切换模型推理后端引擎](../../../../../docs/cn/faq/how_to_change_backend.md)
--- a/examples/vision/keypointdetection/tiny_pose/cpp/pptinypose_infer.cc
+++ b/examples/vision/keypointdetection/tiny_pose/cpp/pptinypose_infer.cc
@@ -0,0 +1,143 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "fastdeploy/vision.h"
+
+#ifdef WIN32
+const char sep = '\\';
+#else
+const char sep = '/';
+#endif
+
+void CpuInfer(const std::string& tinypose_model_dir,
+              const std::string& image_file) {
+  auto tinypose_model_file = tinypose_model_dir + sep + "model.pdmodel";
+  auto tinypose_params_file = tinypose_model_dir + sep + "model.pdiparams";
+  auto tinypose_config_file = tinypose_model_dir + sep + "infer_cfg.yml";
+  auto tinypose_model = fastdeploy::vision::keypointdetection::PPTinyPose(
+      tinypose_model_file, tinypose_params_file, tinypose_config_file);
+  if (!tinypose_model.Initialized()) {
+    std::cerr << "TinyPose Model Failed to initialize." << std::endl;
+    return;
+  }
+
+  auto im = cv::imread(image_file);
+  fastdeploy::vision::KeyPointDetectionResult res;
+  if (!tinypose_model.Predict(&im, &res)) {
+    std::cerr << "TinyPose Prediction Failed." << std::endl;
+    return;
+  } else {
+    std::cout << "TinyPose Prediction Done!" << std::endl;
+  }
+  // 输出预测框结果
+  std::cout << res.Str() << std::endl;
+
+  // 可视化预测结果
+  auto tinypose_vis_im =
+      fastdeploy::vision::VisKeypointDetection(im, res, 0.5);
+  cv::imwrite("tinypose_vis_result.jpg", tinypose_vis_im);
+  std::cout << "TinyPose visualized result saved in ./tinypose_vis_result.jpg"
+            << std::endl;
+}
+
+void GpuInfer(const std::string& tinypose_model_dir,
+              const std::string& image_file) {
+  auto option = fastdeploy::RuntimeOption();
+  option.UseGpu();
+
+  auto tinypose_model_file = tinypose_model_dir + sep + "model.pdmodel";
+  auto tinypose_params_file = tinypose_model_dir + sep + "model.pdiparams";
+  auto tinypose_config_file = tinypose_model_dir + sep + "infer_cfg.yml";
+  auto tinypose_model = fastdeploy::vision::keypointdetection::PPTinyPose(
+      tinypose_model_file, tinypose_params_file, tinypose_config_file, option);
+  if (!tinypose_model.Initialized()) {
+    std::cerr << "TinyPose Model Failed to initialize." << std::endl;
+    return;
+  }
+
+  auto im = cv::imread(image_file);
+  fastdeploy::vision::KeyPointDetectionResult res;
+  if (!tinypose_model.Predict(&im, &res)) {
+    std::cerr << "TinyPose Prediction Failed." << std::endl;
+    return;
+  } else {
+    std::cout << "TinyPose Prediction Done!" << std::endl;
+  }
+  // 输出预测框结果
+  std::cout << res.Str() << std::endl;
+
+  // 可视化预测结果
+  auto tinypose_vis_im =
+      fastdeploy::vision::VisKeypointDetection(im, res, 0.5);
+  cv::imwrite("tinypose_vis_result.jpg", tinypose_vis_im);
+  std::cout << "TinyPose visualized result saved in ./tinypose_vis_result.jpg"
+            << std::endl;
+}
+
+void TrtInfer(const std::string& tinypose_model_dir,
+              const std::string& image_file) {
+  auto tinypose_model_file = tinypose_model_dir + sep + "model.pdmodel";
+  auto tinypose_params_file = tinypose_model_dir + sep + "model.pdiparams";
+  auto tinypose_config_file = tinypose_model_dir + sep + "infer_cfg.yml";
+  auto tinypose_option = fastdeploy::RuntimeOption();
+  tinypose_option.UseGpu();
+  tinypose_option.UseTrtBackend();
+  auto tinypose_model = fastdeploy::vision::keypointdetection::PPTinyPose(
+      tinypose_model_file, tinypose_params_file, tinypose_config_file,
+      tinypose_option);
+  if (!tinypose_model.Initialized()) {
+    std::cerr << "TinyPose Model Failed to initialize." << std::endl;
+    return;
+  }
+
+  auto im = cv::imread(image_file);
+  fastdeploy::vision::KeyPointDetectionResult res;
+  if (!tinypose_model.Predict(&im, &res)) {
+    std::cerr << "TinyPose Prediction Failed." << std::endl;
+    return;
+  } else {
+    std::cout << "TinyPose Prediction Done!" << std::endl;
+  }
+  // 输出预测框结果
+  std::cout << res.Str() << std::endl;
+
+  // 可视化预测结果
+  auto tinypose_vis_im =
+      fastdeploy::vision::VisKeypointDetection(im, res, 0.5);
+  cv::imwrite("tinypose_vis_result.jpg", tinypose_vis_im);
+  std::cout << "TinyPose visualized result saved in ./tinypose_vis_result.jpg"
+            << std::endl;
+}
+
+int main(int argc, char* argv[]) {
+  if (argc < 4) {
+    std::cout << "Usage: infer_demo path/to/pptinypose_model_dir path/to/image "
+                 "run_option, "
+                 "e.g ./infer_model ./pptinypose_model_dir ./test.jpeg 0"
+              << std::endl;
+    std::cout << "The data type of run_option is int, 0: run with cpu; 1: run "
+                 "with gpu; 2: run with gpu and use tensorrt backend."
+              << std::endl;
+    return -1;
+  }
+
+  if (std::atoi(argv[3]) == 0) {
+    CpuInfer(argv[1], argv[2]);
+  } else if (std::atoi(argv[3]) == 1) {
+    GpuInfer(argv[1], argv[2]);
+  } else if (std::atoi(argv[3]) == 2) {
+    TrtInfer(argv[1], argv[2]);
+  }
+  return 0;
+}
--- a/examples/vision/keypointdetection/tiny_pose/python/README.md
+++ b/examples/vision/keypointdetection/tiny_pose/python/README.md
@@ -0,0 +1,79 @@
+# PP-TinyPose Python部署示例
+
+在部署前，需确认以下两个步骤
+
+- 1. 软硬件环境满足要求，参考[FastDeploy环境要求](../../../../../docs/cn/build_and_install/download_prebuilt_libraries.md)  
+- 2. 根据开发环境，下载预编译部署库和samples代码，参考[FastDeploy预编译库](../../../../../docs/cn/build_and_install/download_prebuilt_libraries.md)
+
+本目录下提供`pptinypose_infer.py`快速完成PP-TinyPose在CPU/GPU，以及GPU上通过TensorRT加速部署的`单图单人关键点检测`示例。执行如下脚本即可完成
+
+>> **注意**: PP-Tinypose单模型目前只支持单图单人关键点检测，因此输入的图片应只包含一个人或者进行过裁剪的图像。多人关键点检测请参考[PP-TinyPose Pipeline](../../det_keypoint_unite/python/README.md)
+
+```bash
+#下载部署示例代码
+git clone https://github.com/PaddlePaddle/FastDeploy.git
+cd FastDeploy/examples/vision/keypointdetection/tiny_pose/python
+
+# 下载PP-TinyPose模型文件和测试图片
+wget https://bj.bcebos.com/paddlehub/fastdeploy/PP_TinyPose_256x192_infer.tgz
+tar -xvf PP_TinyPose_256x192_infer.tgz
+wget https://bj.bcebos.com/paddlehub/fastdeploy/hrnet_demo.jpg
+
+# CPU推理
+python pptinypose_infer.py --tinypose_model_dir PP_TinyPose_256x192_infer --image hrnet_demo.jpg --device cpu
+# GPU推理
+python pptinypose_infer.py --tinypose_model_dir PP_TinyPose_256x192_infer --image hrnet_demo.jpg --device gpu
+# GPU上使用TensorRT推理 （注意：TensorRT推理第一次运行，有序列化模型的操作，有一定耗时，需要耐心等待）
+python pptinypose_infer.py --tinypose_model_dir PP_TinyPose_256x192_infer --image hrnet_demo.jpg --device gpu --use_trt True
+```
+
+运行完成可视化结果如下图所示
+<div  align="center">  
+<img src="https://user-images.githubusercontent.com/16222477/196386764-dd51ad56-c410-4c54-9580-643f282f5a83.jpeg", width=359px, height=423px />
+</div>
+
+## PP-TinyPose Python接口
+
+```python
+fd.vision.keypointdetection.PPTinyPose(model_file, params_file, config_file, runtime_option=None, model_format=ModelFormat.PADDLE)
+```
+
+PP-TinyPose模型加载和初始化，其中model_file, params_file以及config_file为训练模型导出的Paddle inference文件，具体请参考其文档说明[模型导出](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.5/deploy/EXPORT_MODEL.md)
+
+**参数**
+
+> * **model_file**(str): 模型文件路径
+> * **params_file**(str): 参数文件路径
+> * **config_file**(str): 推理部署配置文件
+> * **runtime_option**(RuntimeOption): 后端推理配置，默认为None，即采用默认配置
+> * **model_format**(ModelFormat): 模型格式，默认为Paddle格式
+
+### predict函数
+
+> ```python
+> PPTinyPose.predict(input_image)
+> ```
+>
+> 模型预测结口，输入图像直接输出检测结果。
+>
+> **参数**
+>
+> > * **input_image**(np.ndarray): 输入数据，注意需为HWC，BGR格式
+
+> **返回**
+>
+> > 返回`fastdeploy.vision.KeyPointDetectionResult`结构体，结构体说明参考文档[视觉模型预测结果](../../../../../docs/api/vision_results/)
+
+### 类成员属性
+#### 后处理参数
+用户可按照自己的实际需求，修改下列后处理参数，从而影响最终的推理和部署效果
+
+> > * **use_dark**(bool): 是否使用DARK进行后处理[参考论文](https://arxiv.org/abs/1910.06278)
+
+
+## 其它文档
+
+- [PP-TinyPose 模型介绍](..)
+- [PP-TinyPose C++部署](../cpp)
+- [模型预测结果说明](../../../../../docs/api/vision_results/)
+- [如何切换模型推理后端引擎](../../../../../docs/runtime/how_to_change_backend.md)
--- a/examples/vision/keypointdetection/tiny_pose/python/pptinypose_infer.py
+++ b/examples/vision/keypointdetection/tiny_pose/python/pptinypose_infer.py
@@ -0,0 +1,62 @@
+import fastdeploy as fd
+import cv2
+import os
+
+
+def parse_arguments():
+    import argparse
+    import ast
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--tinypose_model_dir",
+        required=True,
+        help="path of paddletinypose model directory")
+    parser.add_argument(
+        "--image", required=True, help="path of test image file.")
+    parser.add_argument(
+        "--device",
+        type=str,
+        default='cpu',
+        help="type of inference device, support 'cpu' or 'gpu'.")
+    parser.add_argument(
+        "--use_trt",
+        type=ast.literal_eval,
+        default=False,
+        help="wether to use tensorrt.")
+    return parser.parse_args()
+
+
+def build_tinypose_option(args):
+    option = fd.RuntimeOption()
+
+    if args.device.lower() == "gpu":
+        option.use_gpu()
+
+    if args.use_trt:
+        option.use_trt_backend()
+        option.set_trt_input_shape("image", [1, 3, 256, 192])
+    return option
+
+
+args = parse_arguments()
+
+tinypose_model_file = os.path.join(args.tinypose_model_dir, "model.pdmodel")
+tinypose_params_file = os.path.join(args.tinypose_model_dir, "model.pdiparams")
+tinypose_config_file = os.path.join(args.tinypose_model_dir, "infer_cfg.yml")
+# 配置runtime，加载模型
+runtime_option = build_tinypose_option(args)
+tinypose_model = fd.vision.keypointdetection.PPTinyPose(
+    tinypose_model_file,
+    tinypose_params_file,
+    tinypose_config_file,
+    runtime_option=runtime_option)
+# 预测图片检测结果
+im = cv2.imread(args.image)
+tinypose_result = tinypose_model.predict(im)
+print("Paddle TinyPose Result:\n", tinypose_result)
+
+# 预测结果可视化
+vis_im = fd.vision.vis_keypoint_detection(
+    im, tinypose_result, conf_threshold=0.5)
+cv2.imwrite("visualized_result.jpg", vis_im)
+print("TinyPose visualized result save in ./visualized_result.jpg")