[RKNPU2]support rknpu2 ClasModel #957 (#964)

* [RKNPU2]support rknpu2 ClasModel #957 * [RKNPU2]support rknpu2 ClasModel #957 * [RKNPU2]support rknpu2 add Resnet50_vd example #957 * [RKNPU2]support rknpu2 add Resnet50_vd example #957 * [RKNPU2]support rknpu2, improve doc #957 * [RKNPU2]support rknpu2, improve doc #957 * [RKNPU2]support rknpu2, improve doc #957 * [RKNPU2]support rknpu2, improve doc #957 * [RKNPU2]support rknpu2, improve doc #957 * [RKNPU2]support rknpu2, improve doc #957 * [RKNPU2]support rknpu2, improve doc #957
2025-10-07 01:22:59 +08:00 · 2022-12-28 17:58:18 +08:00
parent 02425bf280
commit 973c746d06
13 changed files with 390 additions and 17 deletions
--- a/docs/cn/faq/rknpu2/rknpu2.md
+++ b/docs/cn/faq/rknpu2/rknpu2.md
@@ -23,6 +23,7 @@ ONNX模型不能直接调用RK芯片中的NPU进行运算，需要把ONNX模型
 | Segmentation     | PP-HumanSegV2Lite | portrait                      | 133/43             |
 | Segmentation     | PP-HumanSegV2Lite | human                         | 133/43             |
 | Face Detection   | SCRFD             | SCRFD-2.5G-kps-640            | 108/42             |
+| Classification   | ResNet            | ResNet50_vd                   | -/92               |

 ## RKNPU2 Backend推理使用教程

--- a/examples/vision/classification/paddleclas/rknpu2/README.md
+++ b/examples/vision/classification/paddleclas/rknpu2/README.md
@@ -0,0 +1,57 @@
+# PaddleClas 模型RKNPU2部署
+
+## 转换模型
+下面以 ResNet50_vd为例子，教大家如何转换分类模型到RKNN模型。
+
+```bash
+# 安装 paddle2onnx
+pip install paddle2onnx
+
+# 下载ResNet50_vd模型文件和测试图片
+wget https://bj.bcebos.com/paddlehub/fastdeploy/ResNet50_vd_infer.tgz
+tar -xvf ResNet50_vd_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 ResNet50_vd_infer  \
+            --model_filename inference.pdmodel \
+            --params_filename inference.pdiparams  \
+            --save_file ResNet50_vd_infer/ResNet50_vd_infer.onnx  \
+            --enable_dev_version True  \
+            --opset_version 12  \
+            --enable_onnx_checker True
+
+# 固定shape，注意这里的inputs得对应netron.app展示的 inputs 的 name，有可能是image 或者 x
+python -m paddle2onnx.optimize --input_model ResNet50_vd_infer/ResNet50_vd_infer.onnx \
+                               --output_model ResNet50_vd_infer/ResNet50_vd_infer.onnx \
+                               --input_shape_dict "{'inputs':[1,3,224,224]}"
+```                               
+
+ ### 编写模型导出配置文件
+以转化RK3588的RKNN模型为例子，我们需要编辑tools/rknpu2/config/ResNet50_vd_infer_rknn.yaml，来转换ONNX模型到RKNN模型。                              
+
+默认的 mean=0, std=1是在内存做normalize，如果你需要在NPU上执行normalize操作，请根据你的模型配置normalize参数，例如:
+```yaml
+model_path: ./ResNet50_vd_infer.onnx
+output_folder: ./
+target_platform: RK3588
+normalize:
+  mean: [[0.485,0.456,0.406]]
+  std: [[0.229,0.224,0.225]]
+outputs: []
+outputs_nodes: []
+do_quantization: False
+dataset: 
+```
+
+
+# ONNX模型转RKNN模型
+```shell
+python tools/rknpu2/export.py \
+        --config_path tools/rknpu2/config/ResNet50_vd_infer_rknn.yaml \
+        --target_platform rk3588
+```
+
+## 其他链接
+- [Cpp部署](./cpp)
+- [Python部署](./python)
+- [视觉模型预测结果](../../../../../docs/api/vision_results/)
--- a/examples/vision/classification/paddleclas/rknpu2/cpp/CMakeLists.txt
+++ b/examples/vision/classification/paddleclas/rknpu2/cpp/CMakeLists.txt
@@ -0,0 +1,37 @@
+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-0.0.3")
+
+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 ppclas_model_dir 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)
--- a/examples/vision/classification/paddleclas/rknpu2/cpp/README.md
+++ b/examples/vision/classification/paddleclas/rknpu2/cpp/README.md
@@ -0,0 +1,78 @@
+# PaddleClas C++部署示例
+
+本目录下用于展示 ResNet50_vd 模型在RKNPU2上的部署，以下的部署过程以 ResNet50_vd 为例子。
+
+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
+├── images  # 存放图片的文件夹
+├── infer.cc
+├── ppclas_model_dir  # 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 ppclas_model_dir
+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-0.0.3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，转换方案: ([ResNet50_vd RKNN模型](../README.md))。
+
+### 准备测试图片至image文件夹
+```bash
+wget https://gitee.com/paddlepaddle/PaddleClas/raw/release/2.4/deploy/images/ImageNet/ILSVRC2012_val_00000010.jpeg
+```
+
+### 编译example
+
+```bash
+cd build
+cmake ..
+make -j8
+make install
+```
+
+## 运行例程
+
+```bash
+cd ./build/install
+./rknpu_test ./ppclas_model_dir ./images/ILSVRC2012_val_00000010.jpeg
+```
+
+## 运行结果展示
+ClassifyResult(
+label_ids: 153, 
+scores: 0.684570, 
+)
+
+## 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，
+DisablePermute(C++)或`disable_permute(Python)，在预处理阶段禁用数据格式的转换。
+
+## 其它文档
+- [ResNet50_vd Python 部署](../python)
+- [模型预测结果说明](../../../../../../docs/api/vision_results/)
+- [转换ResNet50_vd RKNN模型文档](../README.md)
--- a/examples/vision/classification/paddleclas/rknpu2/cpp/infer.cc
+++ b/examples/vision/classification/paddleclas/rknpu2/cpp/infer.cc
@@ -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.
+
+#include "fastdeploy/vision.h"
+
+void RKNPU2Infer(const std::string& model_dir, const std::string& image_file) {
+  auto model_file = model_dir + "/ResNet50_vd_infer_rk3588.rknn";
+  auto params_file = "";
+  auto config_file = model_dir + "/inference_cls.yaml";
+
+  auto option = fastdeploy::RuntimeOption();
+  option.UseRKNPU2();
+
+  auto format = fastdeploy::ModelFormat::RKNN;
+
+  auto model = fastdeploy::vision::classification::PaddleClasModel(
+      model_file, params_file, config_file,option,format);
+  if (!model.Initialized()) {
+    std::cerr << "Failed to initialize." << std::endl;
+    return;
+  }
+  model.GetPreprocessor().DisablePermute();
+  fastdeploy::TimeCounter tc;
+  tc.Start();
+  auto im = cv::imread(image_file);
+  fastdeploy::vision::ClassifyResult res;
+  if (!model.Predict(im, &res)) {
+    std::cerr << "Failed to predict." << std::endl;
+    return;
+  }
+  // print res
+  std::cout << res.Str() << std::endl;
+  tc.End();
+  tc.PrintInfo("PPClas in RKNPU2");
+}
+
+int main(int argc, char* argv[]) {
+  if (argc < 3) {
+    std::cout
+        << "Usage: rknpu_test path/to/model_dir path/to/image run_option, "
+           "e.g ./rknpu_test ./ppclas_model_dir ./images/ILSVRC2012_val_00000010.jpeg"
+        << std::endl;
+    return -1;
+  }
+  RKNPU2Infer(argv[1], argv[2]);
+  return 0;
+}
--- a/examples/vision/classification/paddleclas/rknpu2/python/README.md
+++ b/examples/vision/classification/paddleclas/rknpu2/python/README.md
@@ -0,0 +1,35 @@
+# PaddleClas 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 ResNet50_vd 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/classification/paddleclas/rknpu2/python
+
+# 下载图片
+wget https://gitee.com/paddlepaddle/PaddleClas/raw/release/2.4/deploy/images/ImageNet/ILSVRC2012_val_00000010.jpeg
+
+# 推理
+python3 infer.py --model_file ./ResNet50_vd_infer/ResNet50_vd_infer_rk3588.rknn  --config_file ResNet50_vd_infer/inference_cls.yaml  --image ILSVRC2012_val_00000010.jpeg
+
+# 运行完成后返回结果如下所示
+ClassifyResult(
+label_ids: 153, 
+scores: 0.684570, 
+)
+```
+
+
+## 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，
+DisablePermute(C++)或`disable_permute(Python)，在预处理阶段禁用数据格式的转换。
+
+## 其它文档
+- [ResNet50_vd C++部署](../cpp)
+- [模型预测结果说明](../../../../../../docs/api/vision_results/)
+- [转换ResNet50_vd RKNN模型文档](../README.md)
--- a/examples/vision/classification/paddleclas/rknpu2/python/infer.py
+++ b/examples/vision/classification/paddleclas/rknpu2/python/infer.py
@@ -0,0 +1,50 @@
+# 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 rknn model.")
+    parser.add_argument("--config_file", required=True, help="Path of config.")
+    parser.add_argument(
+        "--image", type=str, required=True, help="Path of test image file.")
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = parse_arguments()
+
+    model_file = args.model_file
+    params_file = ""
+    config_file = args.config_file
+    # 配置runtime，加载模型
+    runtime_option = fd.RuntimeOption()
+    runtime_option.use_rknpu2()
+    model = fd.vision.classification.ResNet50vd(
+        model_file,
+        params_file,
+        config_file,
+        runtime_option=runtime_option,
+        model_format=fd.ModelFormat.RKNN)
+    # 禁用通道转换    
+    model.preprocessor.disable_permute()
+    im = cv2.imread(args.image)
+    result = model.predict(im, topk=1)
+    print(result)
--- a/fastdeploy/vision/classification/ppcls/model.cc
+++ b/fastdeploy/vision/classification/ppcls/model.cc
@@ -32,9 +32,10 @@ PaddleClasModel::PaddleClasModel(const std::string& model_file,
    valid_ascend_backends = {Backend::LITE};
    valid_kunlunxin_backends = {Backend::LITE};
    valid_ipu_backends = {Backend::PDINFER};
-  } else if (model_format == ModelFormat::ONNX) {
+  } else {
    valid_cpu_backends = {Backend::ORT, Backend::OPENVINO};
    valid_gpu_backends = {Backend::ORT, Backend::TRT};
+    valid_rknpu_backends = {Backend::RKNPU2};
  }
  
  runtime_option = custom_option;
--- a/fastdeploy/vision/classification/ppcls/ppcls_pybind.cc
+++ b/fastdeploy/vision/classification/ppcls/ppcls_pybind.cc
@@ -36,6 +36,12 @@ void BindPaddleClas(pybind11::module& m) {
      })
      .def("use_gpu", [](vision::classification::PaddleClasPreprocessor& self, int gpu_id = -1) {
        self.UseGpu(gpu_id);
+      })
+      .def("disable_normalize", [](vision::classification::PaddleClasPreprocessor& self) {
+        self.DisableNormalize();
+      })
+      .def("disable_permute", [](vision::classification::PaddleClasPreprocessor& self) {
+        self.DisablePermute();
      });

  pybind11::class_<vision::classification::PaddleClasPostprocessor>(
--- a/fastdeploy/vision/classification/ppcls/preprocessor.cc
+++ b/fastdeploy/vision/classification/ppcls/preprocessor.cc
@@ -24,19 +24,19 @@ namespace vision {
 namespace classification {

 PaddleClasPreprocessor::PaddleClasPreprocessor(const std::string& config_file) {
-  FDASSERT(BuildPreprocessPipelineFromConfig(config_file),
+  this->config_file_ = config_file;
+  FDASSERT(BuildPreprocessPipelineFromConfig(),
           "Failed to create PaddleClasPreprocessor.");
  initialized_ = true;
 }

-bool PaddleClasPreprocessor::BuildPreprocessPipelineFromConfig(
-    const std::string& config_file) {
+bool PaddleClasPreprocessor::BuildPreprocessPipelineFromConfig() {
  processors_.clear();
  YAML::Node cfg;
  try {
-    cfg = YAML::LoadFile(config_file);
+    cfg = YAML::LoadFile(config_file_);
  } catch (YAML::BadFile& e) {
-    FDERROR << "Failed to load yaml file " << config_file
+    FDERROR << "Failed to load yaml file " << config_file_
            << ", maybe you should check this file." << std::endl;
    return false;
  }
@@ -57,15 +57,19 @@ bool PaddleClasPreprocessor::BuildPreprocessPipelineFromConfig(
      int height = op.begin()->second["size"].as<int>();
      processors_.push_back(std::make_shared<CenterCrop>(width, height));
    } else if (op_name == "NormalizeImage") {
-      auto mean = op.begin()->second["mean"].as<std::vector<float>>();
-      auto std = op.begin()->second["std"].as<std::vector<float>>();
-      auto scale = op.begin()->second["scale"].as<float>();
-      FDASSERT((scale - 0.00392157) < 1e-06 && (scale - 0.00392157) > -1e-06,
-               "Only support scale in Normalize be 0.00392157, means the pixel "
-               "is in range of [0, 255].");
-      processors_.push_back(std::make_shared<Normalize>(mean, std));
+      if (!disable_normalize) {
+        auto mean = op.begin()->second["mean"].as<std::vector<float>>();
+        auto std = op.begin()->second["std"].as<std::vector<float>>();
+        auto scale = op.begin()->second["scale"].as<float>();
+        FDASSERT((scale - 0.00392157) < 1e-06 && (scale - 0.00392157) > -1e-06,
+                "Only support scale in Normalize be 0.00392157, means the pixel "
+                "is in range of [0, 255].");
+        processors_.push_back(std::make_shared<Normalize>(mean, std));
+      }
    } else if (op_name == "ToCHWImage") {
-      processors_.push_back(std::make_shared<HWC2CHW>());
+      if (!disable_permute) {
+        processors_.push_back(std::make_shared<HWC2CHW>());
+      }
    } else {
      FDERROR << "Unexcepted preprocess operator: " << op_name << "."
              << std::endl;
@@ -78,6 +82,21 @@ bool PaddleClasPreprocessor::BuildPreprocessPipelineFromConfig(
  return true;
 }

+void PaddleClasPreprocessor::DisableNormalize() {
+  this->disable_normalize = true;
+  // the DisableNormalize function will be invalid if the configuration file is loaded during preprocessing
+  if (!BuildPreprocessPipelineFromConfig()) {
+    FDERROR << "Failed to build preprocess pipeline from configuration file." << std::endl;
+  }
+}
+void PaddleClasPreprocessor::DisablePermute() {
+  this->disable_permute = true;
+  // the DisablePermute function will be invalid if the configuration file is loaded during preprocessing
+  if (!BuildPreprocessPipelineFromConfig()) {
+    FDERROR << "Failed to build preprocess pipeline from configuration file." << std::endl;
+  }
+}
+
 void PaddleClasPreprocessor::UseGpu(int gpu_id) {
 #ifdef WITH_GPU
  use_cuda_ = true;
--- a/fastdeploy/vision/classification/ppcls/preprocessor.h
+++ b/fastdeploy/vision/classification/ppcls/preprocessor.h
@@ -46,13 +46,24 @@ class FASTDEPLOY_DECL PaddleClasPreprocessor {

  bool WithGpu() { return use_cuda_; }

+  /// This function will disable normalize in preprocessing step.
+  void DisableNormalize();
+  /// This function will disable hwc2chw in preprocessing step.
+  void DisablePermute();
+
 private:
-  bool BuildPreprocessPipelineFromConfig(const std::string& config_file);
+  bool BuildPreprocessPipelineFromConfig();
  std::vector<std::shared_ptr<Processor>> processors_;
  bool initialized_ = false;
  bool use_cuda_ = false;
  // GPU device id
  int device_id_ = -1;
+  // for recording the switch of hwc2chw
+  bool disable_permute = false;
+  // for recording the switch of normalize
+  bool disable_normalize = false;
+  // read config file
+  std::string config_file_;
 };

 }  // namespace classification
--- a/python/fastdeploy/vision/classification/ppcls/init.py
+++ b/python/fastdeploy/vision/classification/ppcls/init.py
@@ -42,6 +42,18 @@ class PaddleClasPreprocessor:
        """
        return self._preprocessor.use_gpu(gpu_id)

+    def disable_normalize(self):
+        """
+        This function will disable normalize in preprocessing step.
+        """
+        self._preprocessor.disable_normalize()
+
+    def disable_permute(self):
+        """
+        This function will disable hwc2chw in preprocessing step.
+        """
+        self._preprocessor.disable_permute()      
+

 class PaddleClasPostprocessor:
    def __init__(self, topk=1):
@@ -78,8 +90,6 @@ class PaddleClasModel(FastDeployModel):
        """

        super(PaddleClasModel, self).__init__(runtime_option)
-
-        assert model_format == ModelFormat.PADDLE, "PaddleClasModel only support model format of ModelFormat.PADDLE now."
        self._model = C.vision.classification.PaddleClasModel(
            model_file, params_file, config_file, self._runtime_option,
            model_format)
--- a/tools/rknpu2/config/ResNet50_vd_infer_rknn.yaml
+++ b/tools/rknpu2/config/ResNet50_vd_infer_rknn.yaml
@@ -0,0 +1,10 @@
+model_path: ./ResNet50_vd_infer.onnx
+output_folder: ./
+target_platform: RK3588
+normalize:
+  mean: [[0, 0, 0]]
+  std: [[1, 1, 1]]
+outputs: []
+outputs_nodes: []
+do_quantization: False
+dataset: