Merge pull request PaddlePaddle#1339 from Zheng-Bicheng/tiny_pose

leyond123 · Feb 16, 2023 · e3a7ab4 · e3a7ab4
2 parents ee85a3c + b829a4e
commit e3a7ab4
Show file tree

Hide file tree

Showing 12 changed files with 413 additions and 10 deletions.
diff --git a/examples/vision/keypointdetection/tiny_pose/rknpu2/README.md b/examples/vision/keypointdetection/tiny_pose/rknpu2/README.md
@@ -0,0 +1,55 @@
+[English](README.md) | 简体中文
+# PP-TinyPose RKNPU2部署示例
+
+## 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。
+
+## 模型转换example
+
+### Paddle模型转换为ONNX模型
+
+由于Rockchip提供的rknn-toolkit2工具暂时不支持Paddle模型直接导出为RKNN模型，因此需要先将Paddle模型导出为ONNX模型，再将ONNX模型转为RKNN模型。
+
+```bash
+# 下载Paddle静态图模型并解压
+wget https://bj.bcebos.com/paddlehub/fastdeploy/PP_TinyPose_256x192_infer.tgz
+tar -xvf PP_TinyPose_256x192_infer.tgz
+
+# 静态图转ONNX模型，注意，这里的save_file请和压缩包名对齐
+paddle2onnx --model_dir PP_TinyPose_256x192_infer \
+            --model_filename model.pdmodel \
+            --params_filename model.pdiparams \
+            --save_file PP_TinyPose_256x192_infer/PP_TinyPose_256x192_infer.onnx \
+            --enable_dev_version True
+
+# 固定shape
+python -m paddle2onnx.optimize --input_model PP_TinyPose_256x192_infer/PP_TinyPose_256x192_infer.onnx \
+                                --output_model PP_TinyPose_256x192_infer/PP_TinyPose_256x192_infer.onnx \
+                                --input_shape_dict "{'image':[1,3,256,192]}"
+```
+
+### ONNX模型转RKNN模型
+
+为了方便大家使用，我们提供了python脚本，通过我们预配置的config文件，你将能够快速地转换ONNX模型到RKNN模型
+
+```bash
+python tools/rknpu2/export.py --config_path tools/rknpu2/config/PP_TinyPose_256x192_unquantized.yaml \
+                              --target_platform rk3588
+```
+
+## 详细部署文档
+
+- [模型详细介绍](../README_CN.md)
+- [Python部署](./python)
+- [C++部署](./cpp)
diff --git a/examples/vision/keypointdetection/tiny_pose/rknpu2/cpp/CMakeLists.txt b/examples/vision/keypointdetection/tiny_pose/rknpu2/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 ${FASTDEPLOY_LIBS})
diff --git a/examples/vision/keypointdetection/tiny_pose/rknpu2/cpp/README.md b/examples/vision/keypointdetection/tiny_pose/rknpu2/cpp/README.md
@@ -0,0 +1,83 @@
+[English](README.md) | 简体中文
+# PP-TinyPose C++部署示例
+
+本目录下提供`pptinypose_infer.cc`快速完成PP-TinyPose通过NPU加速部署的`单图单人关键点检测`示例
+>> **注意**: PP-Tinypose单模型目前只支持单图单人关键点检测，因此输入的图片应只包含一个人或者进行过裁剪的图像。多人关键点检测请参考[PP-TinyPose Pipeline](../../../det_keypoint_unite/cpp/README.md)
+
+在部署前，需确认以下两个步骤
+
+- 1. 软硬件环境满足要求，参考[FastDeploy环境要求](../../../../../../docs/cn/build_and_install/download_prebuilt_libraries.md)  
+- 2. 根据开发环境，下载预编译部署库和samples代码，参考[FastDeploy预编译库](../../../../../../docs/cn/build_and_install/download_prebuilt_libraries.md)
+
+
+以Linux上推理为例，在本目录执行如下命令即可完成编译测试，支持此模型需保证FastDeploy版本1.0.3以上(x.x.x>=1.0.3)
+
+```bash
+mkdir build
+cd build
+
+cmake .. -DFASTDEPLOY_INSTALL_DIR=${PWD}/fastdeploy-linux-x64-x.x.x
+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
+
+
+# NPU推理
+sudo ./infer_tinypose_demo ./PP_TinyPose_256x192_infer ./hrnet_demo.jpg
+```
+
+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)
diff --git a/examples/vision/keypointdetection/tiny_pose/rknpu2/cpp/pptinypose_infer.cc b/examples/vision/keypointdetection/tiny_pose/rknpu2/cpp/pptinypose_infer.cc
@@ -0,0 +1,66 @@
+// 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"
+
+void RKNPU2Infer(const std::string& tinypose_model_dir,
+                 const std::string& image_file) {
+  auto tinypose_model_file =
+      tinypose_model_dir + "/PP_TinyPose_256x192_infer_rk3588_unquantized.rknn";
+  std::cout << tinypose_model_file << std::endl;
+  auto tinypose_params_file = "";
+  auto tinypose_config_file = tinypose_model_dir + "/infer_cfg.yml";
+  auto option = fastdeploy::RuntimeOption();
+  option.UseRKNPU2();
+  auto tinypose_model = fastdeploy::vision::keypointdetection::PPTinyPose(
+      tinypose_model_file, tinypose_params_file, tinypose_config_file, option,
+      fastdeploy::RKNN);
+
+  if (!tinypose_model.Initialized()) {
+    std::cerr << "TinyPose Model Failed to initialize." << std::endl;
+    return;
+  }
+
+  tinypose_model.DisablePermute();
+  tinypose_model.DisableNormalize();
+
+  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;
+  }
+
+  std::cout << res.Str() << std::endl;
+
+  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 < 3) {
+    std::cout << "Usage: infer_demo path/to/pptinypose_model_dir path/to/image "
+                 "run_option, "
+                 "e.g ./infer_model ./pptinypose_model_dir ./test.jpeg"
+              << std::endl;
+    return -1;
+  }
+
+  RKNPU2Infer(argv[1], argv[2]);
+  return 0;
+}
diff --git a/examples/vision/keypointdetection/tiny_pose/rknpu2/python/README_CN.md b/examples/vision/keypointdetection/tiny_pose/rknpu2/python/README_CN.md
@@ -0,0 +1,70 @@
+[English](README.md) | 简体中文
+# 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NPU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PP-TinyPose模型文件和测试图片
+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
+```
+
+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/cn/faq/how_to_change_backend.md)
diff --git a/examples/vision/keypointdetection/tiny_pose/rknpu2/python/pptinypose_infer.py b/examples/vision/keypointdetection/tiny_pose/rknpu2/python/pptinypose_infer.py
@@ -0,0 +1,50 @@
+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.")
+    return parser.parse_args()
+
+
+def build_tinypose_option(args):
+    option = fd.RuntimeOption()
+    option.use_rknpu()
+    return option
+
+
+args = parse_arguments()
+
+tinypose_model_file = os.path.join(args.tinypose_model_dir, "PP_TinyPose_256x192_infer_rk3588_unquantized.rknn")
+tinypose_params_file = os.path.join(args.tinypose_model_dir, "")
+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,
+    model_format=fd.ModelFormat.RKNN)
+tinypose_model.disable_normalize()
+tinypose_model.disable_permute()
+
+# 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")
diff --git a/examples/vision/segmentation/paddleseg/sophgo/python/README.md b/examples/vision/segmentation/paddleseg/sophgo/python/README.md
@@ -16,7 +16,7 @@ cd path/to/paddleseg/sophgo/python
 wget https://paddleseg.bj.bcebos.com/dygraph/demo/cityscapes_demo.png
 
 # PaddleSeg模型转换为bmodel模型
-将Paddle模型转换为SOPHGO bmodel模型，转换步骤参考[文档](../README_CN.md#将paddleseg推理模型转换为bmodel模型步骤)
+将Paddle模型转换为SOPHGO bmodel模型，转换步骤参考[文档](../README.md#将paddleseg推理模型转换为bmodel模型步骤)
 
 # 推理
 python3 infer.py --model_file ./bmodel/pp_liteseg_1684x_f32.bmodel --config_file ./bmodel/deploy.yaml --image cityscapes_demo.png