[Model] 新增scrfd rknpu2代码 (PaddlePaddle#504)

* * 新增scrfd rknpu2代码 * * 新增scrfd python代码 * 修正文档 * 修正文档以及部分错误 * 修改文档 * 修复部分错误 * 修复部分错误 * 修复部分错误 * scrfd更新代码 * scrfd更新代码
Xiue233 · Nov 10, 2022 · ca0de42 · ca0de42
1 parent 3319a5b
commit ca0de42
Show file tree

Hide file tree

Showing 12 changed files with 395 additions and 11 deletions.
diff --git a/examples/vision/facedet/scrfd/rknpu2/README.md b/examples/vision/facedet/scrfd/rknpu2/README.md
@@ -0,0 +1,73 @@
+# SCRFD RKNPU2部署模型
+
+
+- [SCRFD](https://github.com/deepinsight/insightface/tree/17cdeab12a35efcebc2660453a8cbeae96e20950)
+    - （1）[官方库](https://github.com/deepinsight/insightface/)中提供的*.pt通过[导出ONNX模型](#导出ONNX模型)操作后，可进行部署；
+    - （2）开发者基于自己数据训练的SCRFD模型，可按照[导出ONNX模型](#导出ONNX模型)后，完成部署。
+
+## 下载预训练ONNX模型
+
+为了方便开发者的测试，下面提供了SCRFD导出的各系列模型，开发者可直接下载使用。（下表中模型的精度来源于源官方库）
+| 模型                                                               | 大小    | 精度    |
+|:---------------------------------------------------------------- |:----- |:----- |
+| [SCRFD-500M-kps-160](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_500m_bnkps_shape160x160.onnx) | 2.5MB | - |
+| [SCRFD-500M-160](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_500m_shape160x160.onnx) | 2.2MB | - |
+| [SCRFD-500M-kps-320](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_500m_bnkps_shape320x320.onnx) | 2.5MB | - |
+| [SCRFD-500M-320](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_500m_shape320x320.onnx) | 2.2MB | - |
+| [SCRFD-500M-kps-640](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_500m_bnkps_shape640x640.onnx) | 2.5MB | 90.97% |
+| [SCRFD-500M-640](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_500m_shape640x640.onnx) | 2.2MB | 90.57% |
+| [SCRFD-1G-160](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_1g_shape160x160.onnx ) | 2.5MB | - |
+| [SCRFD-1G-320](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_1g_shape320x320.onnx) | 2.5MB | - |
+| [SCRFD-1G-640](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_1g_shape640x640.onnx) | 2.5MB | 92.38% |
+| [SCRFD-2.5G-kps-160](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_2.5g_bnkps_shape160x160.onnx) | 3.2MB | - |
+| [SCRFD-2.5G-160](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_2.5g_shape160x160.onnx) | 2.6MB | - |
+| [SCRFD-2.5G-kps-320](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_2.5g_bnkps_shape320x320.onnx) | 3.2MB | - |
+| [SCRFD-2.5G-320](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_2.5g_shape320x320.onnx) | 2.6MB | - |
+| [SCRFD-2.5G-kps-640](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_2.5g_bnkps_shape640x640.onnx) | 3.2MB | 93.8% |
+| [SCRFD-2.5G-640](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_2.5g_shape640x640.onnx) | 2.6MB | 93.78% |
+| [SCRFD-10G-kps-320](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_10g_bnkps_shape320x320.onnx) | 17MB | - |
+| [SCRFD-10G-320](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_10g_shape320x320.onnx) | 15MB | - |
+| [SCRFD-10G-kps-640](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_10g_bnkps_shape640x640.onnx) | 17MB | 95.4% |
+| [SCRFD-10G-640](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_10g_shape640x640.onnx) | 15MB | 95.16% |
+| [SCRFD-10G-kps-1280](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_10g_bnkps_shape1280x1280.onnx) | 17MB | - |
+| [SCRFD-10G-1280](https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_10g_shape1280x1280.onnx) | 15MB | - |
+
+## 导出ONNX模型
+
+  ```bash
+  #下载scrfd模型文件
+  e.g. download from  https://onedrive.live.com/?authkey=%21ABbFJx2JMhNjhNA&id=4A83B6B633B029CC%215542&cid=4A83B6B633B029CC
+
+  # 安装官方库配置环境，此版本导出环境为：
+  - 手动配置环境
+    torch==1.8.0
+    mmcv==1.3.5
+    mmdet==2.7.0
+
+  - 通过docker配置
+    docker pull qyjdefdocker/onnx-scrfd-converter:v0.3
+
+  # 导出onnx格式文件
+  - 手动生成
+    python tools/scrfd2onnx.py configs/scrfd/scrfd_500m.py weights/scrfd_500m.pth --shape 640 --input-img face-xxx.jpg
+
+  - docker
+    docker的onnx目录中已有生成好的onnx文件
+
+  ```
+
+## ONNX模型转换RKNN模型
+```bash
+wget  https://bj.bcebos.com/paddlehub/fastdeploy/scrfd_500m_bnkps_shape640x640.onnx
+python tools/rknpu2/export.py --config_path tools/rknpu2/config/RK3588/scrfd.yaml
+```
+
+## 详细部署文档
+
+- [Python部署](python/README.md)
+- [C++部署](cpp/README.md)
+
+
+## 版本说明
+
+- 本版本文档和代码基于[SCRFD CommitID:17cdeab](https://github.com/deepinsight/insightface/tree/17cdeab12a35efcebc2660453a8cbeae96e20950) 编写
diff --git a/examples/vision/facedet/scrfd/rknpu2/cpp/CMakeLists.txt b/examples/vision/facedet/scrfd/rknpu2/cpp/CMakeLists.txt
@@ -0,0 +1,36 @@
+CMAKE_MINIMUM_REQUIRED(VERSION 3.10)
+project(rknpu_test)
+
+set(CMAKE_CXX_STANDARD 14)
+
+# 指定下载解压后的fastdeploy库路径
+set(FASTDEPLOY_INSTALL_DIR "thirdpartys/fastdeploy-0.7.0")
+
+include(${FASTDEPLOY_INSTALL_DIR}/FastDeployConfig.cmake)
+include_directories(${FastDeploy_INCLUDE_DIRS})
+add_executable(rknpu_test infer.cc)
+target_link_libraries(rknpu_test
+        ${FastDeploy_LIBS}
+        )
+
+set(CMAKE_INSTALL_PREFIX ${CMAKE_SOURCE_DIR}/build/install)
+
+install(TARGETS rknpu_test DESTINATION ./)
+
+install(DIRECTORY model DESTINATION ./)
+install(DIRECTORY images DESTINATION ./)
+
+file(GLOB FASTDEPLOY_LIBS ${FASTDEPLOY_INSTALL_DIR}/lib/*)
+message("${FASTDEPLOY_LIBS}")
+install(PROGRAMS ${FASTDEPLOY_LIBS} DESTINATION lib)
+
+file(GLOB ONNXRUNTIME_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/onnxruntime/lib/*)
+install(PROGRAMS ${ONNXRUNTIME_LIBS} DESTINATION lib)
+
+install(DIRECTORY ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/opencv/lib DESTINATION ./)
+
+file(GLOB PADDLETOONNX_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/paddle2onnx/lib/*)
+install(PROGRAMS ${PADDLETOONNX_LIBS} DESTINATION lib)
+
+file(GLOB RKNPU2_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/rknpu2_runtime/RK3588/lib/*)
+install(PROGRAMS ${RKNPU2_LIBS} DESTINATION lib)
diff --git a/examples/vision/facedet/scrfd/rknpu2/cpp/README.md b/examples/vision/facedet/scrfd/rknpu2/cpp/README.md
@@ -0,0 +1,72 @@
+# SCRFD C++部署示例
+
+本目录下提供`infer.cc`快速完成SCRFD在NPU加速部署的示例。
+
+在部署前，需确认以下两个步骤:
+
+1. 软硬件环境满足要求
+2. 根据开发环境，下载预编译部署库或者从头编译FastDeploy仓库
+
+以上步骤请参考[RK2代NPU部署库编译](../../../../../../docs/cn/build_and_install/rknpu2.md)实现
+
+## 生成基本目录文件
+
+该例程由以下几个部分组成
+```text
+.
+├── CMakeLists.txt
+├── build  # 编译文件夹
+├── image  # 存放图片的文件夹
+├── infer_cpu_npu.cc
+├── infer_cpu_npu.h
+├── main.cc
+├── model  # 存放模型文件的文件夹
+└── thirdpartys  # 存放sdk的文件夹
+```
+
+首先需要先生成目录结构
+```bash
+mkdir build
+mkdir images
+mkdir model
+mkdir thirdpartys
+```
+
+## 编译
+
+### 编译并拷贝SDK到thirdpartys文件夹
+
+请参考[RK2代NPU部署库编译](../../../../../../docs/cn/build_and_install/rknpu2.md)仓库编译SDK，编译完成后，将在build目录下生成
+fastdeploy-0.7.0目录，请移动它至thirdpartys目录下.
+
+### 拷贝模型文件，以及配置文件至model文件夹
+在Paddle动态图模型 -> Paddle静态图模型 -> ONNX模型的过程中，将生成ONNX文件以及对应的yaml配置文件，请将配置文件存放到model文件夹内。
+转换为RKNN后的模型文件也需要拷贝至model。
+
+### 准备测试图片至image文件夹
+```bash
+wget https://raw.githubusercontent.com/DefTruth/lite.ai.toolkit/main/examples/lite/resources/test_lite_face_detector_3.jpg
+cp test_lite_face_detector_3.jpg ./images
+```
+
+### 编译example
+
+```bash
+cd build
+cmake ..
+make -j8
+make install
+```
+## 运行例程
+
+```bash
+cd ./build/install
+./rknpu_test
+```
+运行完成可视化结果如下图所示
+
+<img width="640" src="https://user-images.githubusercontent.com/67993288/184301789-1981d065-208f-4a6b-857c-9a0f9a63e0b1.jpg">
+
+- [模型介绍](../../README.md)
+- [Python部署](../python/README.md)
+- [视觉模型预测结果](../../../../../../docs/api/vision_results/README.md)
diff --git a/examples/vision/facedet/scrfd/rknpu2/cpp/infer.cc b/examples/vision/facedet/scrfd/rknpu2/cpp/infer.cc
@@ -0,0 +1,79 @@
+#include <iostream>
+#include <string>
+#include "fastdeploy/vision.h"
+
+void InferScrfd(const std::string& device = "cpu");
+
+int main() {
+  InferScrfd("npu");
+  return 0;
+}
+
+fastdeploy::RuntimeOption GetOption(const std::string& device) {
+  auto option = fastdeploy::RuntimeOption();
+  if (device == "npu") {
+    option.UseRKNPU2();
+  } else {
+    option.UseCpu();
+  }
+  return option;
+}
+
+fastdeploy::ModelFormat GetFormat(const std::string& device) {
+  auto format = fastdeploy::ModelFormat::ONNX;
+  if (device == "npu") {
+    format = fastdeploy::ModelFormat::RKNN;
+  } else {
+    format = fastdeploy::ModelFormat::ONNX;
+  }
+  return format;
+}
+
+std::string GetModelPath(std::string& model_path, const std::string& device) {
+  if (device == "npu") {
+    model_path += "rknn";
+  } else {
+    model_path += "onnx";
+  }
+  return model_path;
+}
+
+void InferScrfd(const std::string& device) {
+  std::string model_file =
+      "./model/scrfd_500m_bnkps_shape640x640_rk3588.";
+  std::string params_file;
+
+  fastdeploy::RuntimeOption option = GetOption(device);
+  fastdeploy::ModelFormat format = GetFormat(device);
+  model_file = GetModelPath(model_file, device);
+  auto model = fastdeploy::vision::facedet::SCRFD(
+      model_file, params_file, option, format);
+
+  if (!model.Initialized()) {
+    std::cerr << "Failed to initialize." << std::endl;
+    return;
+  }
+  auto image_file =
+      "./images/test_lite_face_detector_3.jpg";
+  auto im = cv::imread(image_file);
+
+  if (device == "npu") {
+    model.DisableNormalizeAndPermute();
+  }
+
+  fastdeploy::vision::FaceDetectionResult res;
+  clock_t start = clock();
+  if (!model.Predict(&im, &res)) {
+    std::cerr << "Failed to predict." << std::endl;
+    return;
+  }
+  clock_t end = clock();
+  auto dur = static_cast<double>(end - start);
+  printf("InferScrfd use time:%f\n",
+         (dur / CLOCKS_PER_SEC));
+
+  std::cout << res.Str() << std::endl;
+  auto vis_im = fastdeploy::vision::Visualize::VisFaceDetection(im, res);
+  cv::imwrite("vis_result.jpg", vis_im);
+  std::cout << "Visualized result saved in ./vis_result.jpg" << std::endl;
+}
diff --git a/examples/vision/facedet/scrfd/rknpu2/python/README.md b/examples/vision/facedet/scrfd/rknpu2/python/README.md
@@ -0,0 +1,32 @@
+# SCRFD 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/rknpu2.md)
+
+
+本目录下提供`infer.py`快速完成SCRFD在RKNPU上部署的示例。执行如下脚本即可完成
+
+```bash
+# 下载部署示例代码
+git clone https://github.com/PaddlePaddle/FastDeploy.git
+cd FastDeploy/examples/vision/facedet/scrfd/rknpu2/python
+
+# 下载图片
+wget https://raw.githubusercontent.com/DefTruth/lite.ai.toolkit/main/examples/lite/resources/test_lite_face_detector_3.jpg
+
+# 推理
+python3 infer.py --model_file ./scrfd_500m_bnkps_shape640x640_rk3588.rknn \
+                 --image test_lite_face_detector_3.jpg
+```
+
+
+## 注意事项
+RKNPU上对模型的输入要求是使用NHWC格式，且图片归一化操作会在转RKNN模型时，内嵌到模型中，因此我们在使用FastDeploy部署时，
+需要先调用DisableNormalizePermute(C++)或`disable_normalize_permute(Python)，在预处理阶段禁用归一化以及数据格式的转换。
+## 其它文档
+
+- [SCRFD 模型介绍](../README.md)
+- [SCRFD C++部署](../cpp/README.md)
+- [模型预测结果说明](../../../../../../docs/api/vision_results/README.md)
+- [转换SCRFD RKNN模型文档](../README.md)
diff --git a/examples/vision/facedet/scrfd/rknpu2/python/infer.py b/examples/vision/facedet/scrfd/rknpu2/python/infer.py
@@ -0,0 +1,58 @@
+# 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.
+import fastdeploy as fd
+import cv2
+import os
+
+
+def parse_arguments():
+    import argparse
+    import ast
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model_file", required=True, help="Path of FaceDet model.")
+    parser.add_argument(
+        "--image", type=str, required=True, help="Path of test image file.")
+    return parser.parse_args()
+
+
+def build_option(args):
+    option = fd.RuntimeOption()
+    option.use_rknpu2()
+    return option
+
+
+args = parse_arguments()
+
+# 配置runtime，加载模型
+runtime_option = build_option(args)
+model_file = args.model_file
+params_file = ""
+model = fd.vision.facedet.SCRFD(
+    model_file,
+    params_file,
+    runtime_option=runtime_option,
+    model_format=fd.ModelFormat.RKNN)
+
+model.disable_normalize_and_permute()
+
+# 预测图片分割结果
+im = cv2.imread(args.image)
+result = model.predict(im.copy())
+print(result)
+
+# 可视化结果
+vis_im = fd.vision.vis_face_detection(im, result)
+cv2.imwrite("visualized_result.jpg", vis_im)
+print("Visualized result save in ./visualized_result.jpg")