Modify PPMatting backend and docs (PaddlePaddle#182)

* first commit for yolov7 * pybind for yolov7 * CPP README.md * CPP README.md * modified yolov7.cc * README.md * python file modify * delete license in fastdeploy/ * repush the conflict part * README.md modified * README.md modified * file path modified * file path modified * file path modified * file path modified * file path modified * README modified * README modified * move some helpers to private * add examples for yolov7 * api.md modified * api.md modified * api.md modified * YOLOv7 * yolov7 release link * yolov7 release link * yolov7 release link * copyright * change some helpers to private * change variables to const and fix documents. * gitignore * Transfer some funtions to private member of class * Transfer some funtions to private member of class * Merge from develop (#9) * Fix compile problem in different python version (#26) * fix some usage problem in linux * Fix compile problem Co-authored-by: root <[email protected]> * Add PaddleDetetion/PPYOLOE model support (#22) * add ppdet/ppyoloe * Add demo code and documents * add convert processor to vision (#27) * update .gitignore * Added checking for cmake include dir * fixed missing trt_backend option bug when init from trt * remove un-need data layout and add pre-check for dtype * changed RGB2BRG to BGR2RGB in ppcls model * add model_zoo yolov6 c++/python demo * fixed CMakeLists.txt typos * update yolov6 cpp/README.md * add yolox c++/pybind and model_zoo demo * move some helpers to private * fixed CMakeLists.txt typos * add normalize with alpha and beta * add version notes for yolov5/yolov6/yolox * add copyright to yolov5.cc * revert normalize * fixed some bugs in yolox * fixed examples/CMakeLists.txt to avoid conflicts * add convert processor to vision * format examples/CMakeLists summary * Fix bug while the inference result is empty with YOLOv5 (#29) * Add multi-label function for yolov5 * Update README.md Update doc * Update fastdeploy_runtime.cc fix variable option.trt_max_shape wrong name * Update runtime_option.md Update resnet model dynamic shape setting name from images to x * Fix bug when inference result boxes are empty * Delete detection.py Co-authored-by: Jason <[email protected]> Co-authored-by: root <[email protected]> Co-authored-by: DefTruth <[email protected]> Co-authored-by: huangjianhui <[email protected]> * first commit for yolor * for merge * Develop (#11) * Fix compile problem in different python version (#26) * fix some usage problem in linux * Fix compile problem Co-authored-by: root <[email protected]> * Add PaddleDetetion/PPYOLOE model support (#22) * add ppdet/ppyoloe * Add demo code and documents * add convert processor to vision (#27) * update .gitignore * Added checking for cmake include dir * fixed missing trt_backend option bug when init from trt * remove un-need data layout and add pre-check for dtype * changed RGB2BRG to BGR2RGB in ppcls model * add model_zoo yolov6 c++/python demo * fixed CMakeLists.txt typos * update yolov6 cpp/README.md * add yolox c++/pybind and model_zoo demo * move some helpers to private * fixed CMakeLists.txt typos * add normalize with alpha and beta * add version notes for yolov5/yolov6/yolox * add copyright to yolov5.cc * revert normalize * fixed some bugs in yolox * fixed examples/CMakeLists.txt to avoid conflicts * add convert processor to vision * format examples/CMakeLists summary * Fix bug while the inference result is empty with YOLOv5 (#29) * Add multi-label function for yolov5 * Update README.md Update doc * Update fastdeploy_runtime.cc fix variable option.trt_max_shape wrong name * Update runtime_option.md Update resnet model dynamic shape setting name from images to x * Fix bug when inference result boxes are empty * Delete detection.py Co-authored-by: Jason <[email protected]> Co-authored-by: root <[email protected]> Co-authored-by: DefTruth <[email protected]> Co-authored-by: huangjianhui <[email protected]> * Yolor (#16) * Develop (#11) (#12) * Fix compile problem in different python version (#26) * fix some usage problem in linux * Fix compile problem Co-authored-by: root <[email protected]> * Add PaddleDetetion/PPYOLOE model support (#22) * add ppdet/ppyoloe * Add demo code and documents * add convert processor to vision (#27) * update .gitignore * Added checking for cmake include dir * fixed missing trt_backend option bug when init from trt * remove un-need data layout and add pre-check for dtype * changed RGB2BRG to BGR2RGB in ppcls model * add model_zoo yolov6 c++/python demo * fixed CMakeLists.txt typos * update yolov6 cpp/README.md * add yolox c++/pybind and model_zoo demo * move some helpers to private * fixed CMakeLists.txt typos * add normalize with alpha and beta * add version notes for yolov5/yolov6/yolox * add copyright to yolov5.cc * revert normalize * fixed some bugs in yolox * fixed examples/CMakeLists.txt to avoid conflicts * add convert processor to vision * format examples/CMakeLists summary * Fix bug while the inference result is empty with YOLOv5 (#29) * Add multi-label function for yolov5 * Update README.md Update doc * Update fastdeploy_runtime.cc fix variable option.trt_max_shape wrong name * Update runtime_option.md Update resnet model dynamic shape setting name from images to x * Fix bug when inference result boxes are empty * Delete detection.py Co-authored-by: Jason <[email protected]> Co-authored-by: root <[email protected]> Co-authored-by: DefTruth <[email protected]> Co-authored-by: huangjianhui <[email protected]> Co-authored-by: Jason <[email protected]> Co-authored-by: root <[email protected]> Co-authored-by: DefTruth <[email protected]> Co-authored-by: huangjianhui <[email protected]> * Develop (#13) * Fix compile problem in different python version (#26) * fix some usage problem in linux * Fix compile problem Co-authored-by: root <[email protected]> * Add PaddleDetetion/PPYOLOE model support (#22) * add ppdet/ppyoloe * Add demo code and documents * add convert processor to vision (#27) * update .gitignore * Added checking for cmake include dir * fixed missing trt_backend option bug when init from trt * remove un-need data layout and add pre-check for dtype * changed RGB2BRG to BGR2RGB in ppcls model * add model_zoo yolov6 c++/python demo * fixed CMakeLists.txt typos * update yolov6 cpp/README.md * add yolox c++/pybind and model_zoo demo * move some helpers to private * fixed CMakeLists.txt typos * add normalize with alpha and beta * add version notes for yolov5/yolov6/yolox * add copyright to yolov5.cc * revert normalize * fixed some bugs in yolox * fixed examples/CMakeLists.txt to avoid conflicts * add convert processor to vision * format examples/CMakeLists summary * Fix bug while the inference result is empty with YOLOv5 (#29) * Add multi-label function for yolov5 * Update README.md Update doc * Update fastdeploy_runtime.cc fix variable option.trt_max_shape wrong name * Update runtime_option.md Update resnet model dynamic shape setting name from images to x * Fix bug when inference result boxes are empty * Delete detection.py Co-authored-by: Jason <[email protected]> Co-authored-by: root <[email protected]> Co-authored-by: DefTruth <[email protected]> Co-authored-by: huangjianhui <[email protected]> * documents * documents * documents * documents * documents * documents * documents * documents * documents * documents * documents * documents * Develop (#14) * Fix compile problem in different python version (#26) * fix some usage problem in linux * Fix compile problem Co-authored-by: root <[email protected]> * Add PaddleDetetion/PPYOLOE model support (#22) * add ppdet/ppyoloe * Add demo code and documents * add convert processor to vision (#27) * update .gitignore * Added checking for cmake include dir * fixed missing trt_backend option bug when init from trt * remove un-need data layout and add pre-check for dtype * changed RGB2BRG to BGR2RGB in ppcls model * add model_zoo yolov6 c++/python demo * fixed CMakeLists.txt typos * update yolov6 cpp/README.md * add yolox c++/pybind and model_zoo demo * move some helpers to private * fixed CMakeLists.txt typos * add normalize with alpha and beta * add version notes for yolov5/yolov6/yolox * add copyright to yolov5.cc * revert normalize * fixed some bugs in yolox * fixed examples/CMakeLists.txt to avoid conflicts * add convert processor to vision * format examples/CMakeLists summary * Fix bug while the inference result is empty with YOLOv5 (#29) * Add multi-label function for yolov5 * Update README.md Update doc * Update fastdeploy_runtime.cc fix variable option.trt_max_shape wrong name * Update runtime_option.md Update resnet model dynamic shape setting name from images to x * Fix bug when inference result boxes are empty * Delete detection.py Co-authored-by: root <[email protected]> Co-authored-by: DefTruth <[email protected]> Co-authored-by: huangjianhui <[email protected]> Co-authored-by: Jason <[email protected]> Co-authored-by: root <[email protected]> Co-authored-by: DefTruth <[email protected]> Co-authored-by: huangjianhui <[email protected]> Co-authored-by: Jason <[email protected]> * add is_dynamic for YOLO series (#22) * modify ppmatting backend and docs * modify ppmatting docs * fix the PPMatting size problem * fix LimitShort's log * retrigger ci * modify PPMatting docs * modify the way for dealing with LimitShort Co-authored-by: Jason <[email protected]> Co-authored-by: root <[email protected]> Co-authored-by: DefTruth <[email protected]> Co-authored-by: huangjianhui <[email protected]> Co-authored-by: Jason <[email protected]>
ziqi-jin · Sep 7, 2022 · 7e00c5f · 7e00c5f
1 parent f96fdad
commit 7e00c5f
Show file tree

Hide file tree

Showing 13 changed files with 193 additions and 12 deletions.
diff --git a/csrc/fastdeploy/vision/common/processors/limit_short.h b/csrc/fastdeploy/vision/common/processors/limit_short.h
@@ -34,6 +34,7 @@ class LimitShort : public Processor {
 
   static bool Run(Mat* mat, int max_short = -1, int min_short = -1,
                   ProcLib lib = ProcLib::OPENCV_CPU);
+  int GetMaxShort() { return max_short_; }
 
  private:
   int max_short_;

diff --git a/csrc/fastdeploy/vision/common/processors/resize_by_long.cc b/csrc/fastdeploy/vision/common/processors/resize_by_long.cc
@@ -0,0 +1,78 @@
+// 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/common/processors/resize_by_long.h"
+
+namespace fastdeploy {
+namespace vision {
+
+bool ResizeByLong::CpuRun(Mat* mat) {
+  cv::Mat* im = mat->GetCpuMat();
+  int origin_w = im->cols;
+  int origin_h = im->rows;
+  double scale = GenerateScale(origin_w, origin_h);
+  if (use_scale_) {
+    cv::resize(*im, *im, cv::Size(), scale, scale, interp_);
+  } else {
+    int width = static_cast<int>(round(scale * im->cols));
+    int height = static_cast<int>(round(scale * im->rows));
+    cv::resize(*im, *im, cv::Size(width, height), 0, 0, interp_);
+  }
+  mat->SetWidth(im->cols);
+  mat->SetHeight(im->rows);
+  return true;
+}
+
+#ifdef ENABLE_OPENCV_CUDA
+bool ResizeByLong::GpuRun(Mat* mat) {
+  cv::cuda::GpuMat* im = mat->GetGpuMat();
+  int origin_w = im->cols;
+  int origin_h = im->rows;
+  double scale = GenerateScale(origin_w, origin_h);
+  im->convertTo(*im, CV_32FC(im->channels()));
+  if (use_scale_) {
+    cv::cuda::resize(*im, *im, cv::Size(), scale, scale, interp_);
+  } else {
+    int width = static_cast<int>(round(scale * im->cols));
+    int height = static_cast<int>(round(scale * im->rows));
+    cv::cuda::resize(*im, *im, cv::Size(width, height), 0, 0, interp_);
+  }
+  mat->SetWidth(im->cols);
+  mat->SetHeight(im->rows);
+  return true;
+}
+#endif
+
+double ResizeByLong::GenerateScale(const int origin_w, const int origin_h) {
+  int im_size_max = std::max(origin_w, origin_h);
+  int im_size_min = std::min(origin_w, origin_h);
+  double scale = 1.0f;
+  if (target_size_ == -1) {
+    if (im_size_max > max_size_) {
+      scale = static_cast<double>(max_size_) / static_cast<double>(im_size_max);
+    }
+  } else {
+    scale =
+        static_cast<double>(target_size_) / static_cast<double>(im_size_max);
+  }
+  return scale;
+}
+
+bool ResizeByLong::Run(Mat* mat, int target_size, int interp, bool use_scale,
+                       int max_size, ProcLib lib) {
+  auto r = ResizeByLong(target_size, interp, use_scale, max_size);
+  return r(mat, lib);
+}
+}  // namespace vision
+}  // namespace fastdeploy
diff --git a/csrc/fastdeploy/vision/common/processors/resize_by_long.h b/csrc/fastdeploy/vision/common/processors/resize_by_long.h
@@ -0,0 +1,49 @@
+// 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.
+
+#pragma once
+
+#include "fastdeploy/vision/common/processors/base.h"
+
+namespace fastdeploy {
+namespace vision {
+
+class ResizeByLong : public Processor {
+ public:
+  ResizeByLong(int target_size, int interp = 1, bool use_scale = true,
+               int max_size = -1) {
+    target_size_ = target_size;
+    max_size_ = max_size;
+    interp_ = interp;
+    use_scale_ = use_scale;
+  }
+  bool CpuRun(Mat* mat);
+#ifdef ENABLE_OPENCV_CUDA
+  bool GpuRun(Mat* mat);
+#endif
+  std::string Name() { return "ResizeByLong"; }
+
+  static bool Run(Mat* mat, int target_size, int interp = 1,
+                  bool use_scale = true, int max_size = -1,
+                  ProcLib lib = ProcLib::OPENCV_CPU);
+
+ private:
+  double GenerateScale(const int origin_w, const int origin_h);
+  int target_size_;
+  int max_size_;
+  int interp_;
+  bool use_scale_;
+};
+}  // namespace vision
+}  // namespace fastdeploy
diff --git a/csrc/fastdeploy/vision/common/processors/transform.h b/csrc/fastdeploy/vision/common/processors/transform.h
@@ -24,6 +24,7 @@
 #include "fastdeploy/vision/common/processors/pad.h"
 #include "fastdeploy/vision/common/processors/pad_to_size.h"
 #include "fastdeploy/vision/common/processors/resize.h"
+#include "fastdeploy/vision/common/processors/resize_by_long.h"
 #include "fastdeploy/vision/common/processors/resize_by_short.h"
 #include "fastdeploy/vision/common/processors/resize_to_int_mult.h"
 #include "fastdeploy/vision/common/processors/stride_pad.h"
diff --git a/csrc/fastdeploy/vision/matting/ppmatting/ppmatting.cc b/csrc/fastdeploy/vision/matting/ppmatting/ppmatting.cc
@@ -26,8 +26,8 @@ PPMatting::PPMatting(const std::string& model_file,
                      const RuntimeOption& custom_option,
                      const Frontend& model_format) {
   config_file_ = config_file;
-  valid_cpu_backends = {Backend::PDINFER, Backend::ORT};
-  valid_gpu_backends = {Backend::PDINFER, Backend::ORT, Backend::TRT};
+  valid_cpu_backends = {Backend::ORT, Backend::PDINFER};
+  valid_gpu_backends = {Backend::PDINFER, Backend::TRT};
   runtime_option = custom_option;
   runtime_option.model_format = model_format;
   runtime_option.model_file = model_file;
@@ -74,6 +74,11 @@ bool PPMatting::BuildPreprocessPipelineFromConfig() {
         if (op["min_short"]) {
           min_short = op["min_short"].as<int>();
         }
+        FDINFO << "Detected LimitShort processing step in yaml file, if the "
+                  "model is exported from PaddleSeg, please make sure the "
+                  "input of your model is fixed with a square shape, and "
+                  "greater than or equal to "
+               << max_short << "." << std::endl;
         processors_.push_back(
             std::make_shared<LimitShort>(max_short, min_short));
       } else if (op["type"].as<std::string>() == "ResizeToIntMult") {
@@ -92,6 +97,19 @@ bool PPMatting::BuildPreprocessPipelineFromConfig() {
           std = op["std"].as<std::vector<float>>();
         }
         processors_.push_back(std::make_shared<Normalize>(mean, std));
+      } else if (op["type"].as<std::string>() == "ResizeByLong") {
+        int target_size = op["target_size"].as<int>();
+        processors_.push_back(std::make_shared<ResizeByLong>(target_size));
+      } else if (op["type"].as<std::string>() == "Pad") {
+        // size: (w, h)
+        auto size = op["size"].as<std::vector<int>>();
+        std::vector<float> value = {127.5, 127.5, 127.5};
+        if (op["fill_value"]) {
+          auto value = op["fill_value"].as<std::vector<float>>();
+        }
+        processors_.push_back(std::make_shared<Cast>("float"));
+        processors_.push_back(
+            std::make_shared<PadToSize>(size[1], size[0], value));
       }
     }
     processors_.push_back(std::make_shared<HWC2CHW>());
@@ -102,11 +120,30 @@ bool PPMatting::BuildPreprocessPipelineFromConfig() {
 bool PPMatting::Preprocess(Mat* mat, FDTensor* output,
                            std::map<std::string, std::array<int, 2>>* im_info) {
   for (size_t i = 0; i < processors_.size(); ++i) {
+    if (processors_[i]->Name().compare("LimitShort") == 0) {
+      int input_h = static_cast<int>(mat->Height());
+      int input_w = static_cast<int>(mat->Width());
+      auto processor = dynamic_cast<LimitShort*>(processors_[i].get());
+      int max_short = processor->GetMaxShort();
+      if (runtime_option.backend != Backend::PDINFER) {
+        if (input_w != input_h || input_h < max_short || input_w < max_short) {
+          FDWARNING << "Detected LimitShort processing step in yaml file and "
+                       "the size of input image is Unqualified, Fastdeploy "
+                       "will resize the input image into ("
+                    << max_short << "," << max_short << ")." << std::endl;
+          Resize::Run(mat, max_short, max_short);
+        }
+      }
+    }
     if (!(*(processors_[i].get()))(mat)) {
       FDERROR << "Failed to process image data in " << processors_[i]->Name()
               << "." << std::endl;
       return false;
     }
+    if (processors_[i]->Name().compare("ResizeByLong") == 0) {
+      (*im_info)["resize_by_long"] = {static_cast<int>(mat->Height()),
+                                      static_cast<int>(mat->Width())};
+    }
   }
 
   // Record output shape of preprocessed image
@@ -135,6 +172,7 @@ bool PPMatting::Postprocess(
   // 先获取alpha并resize (使用opencv)
   auto iter_ipt = im_info.find("input_shape");
   auto iter_out = im_info.find("output_shape");
+  auto resize_by_long = im_info.find("resize_by_long");
   FDASSERT(iter_out != im_info.end() && iter_ipt != im_info.end(),
            "Cannot find input_shape or output_shape from im_info.");
   int out_h = iter_out->second[0];
@@ -145,7 +183,17 @@ bool PPMatting::Postprocess(
   // TODO: 需要修改成FDTensor或Mat的运算 现在依赖cv::Mat
   float* alpha_ptr = static_cast<float*>(alpha_tensor.Data());
   cv::Mat alpha_zero_copy_ref(out_h, out_w, CV_32FC1, alpha_ptr);
-  Mat alpha_resized(alpha_zero_copy_ref);  // ref-only, zero copy.
+  cv::Mat cropped_alpha;
+  if (resize_by_long != im_info.end()) {
+    int resize_h = resize_by_long->second[0];
+    int resize_w = resize_by_long->second[1];
+    alpha_zero_copy_ref(cv::Rect(0, 0, resize_w, resize_h))
+        .copyTo(cropped_alpha);
+  } else {
+    cropped_alpha = alpha_zero_copy_ref;
+  }
+  Mat alpha_resized(cropped_alpha);  // ref-only, zero copy.
+
   if ((out_h != ipt_h) || (out_w != ipt_w)) {
     // already allocated a new continuous memory after resize.
     // cv::resize(alpha_resized, alpha_resized, cv::Size(ipt_w, ipt_h));

diff --git a/csrc/fastdeploy/vision/matting/ppmatting/ppmatting.h b/csrc/fastdeploy/vision/matting/ppmatting/ppmatting.h
@@ -27,7 +27,7 @@ class FASTDEPLOY_DECL PPMatting : public FastDeployModel {
             const RuntimeOption& custom_option = RuntimeOption(),
             const Frontend& model_format = Frontend::PADDLE);
 
-  std::string ModelName() const { return "PaddleMat"; }
+  std::string ModelName() const { return "PaddleMatting"; }
 
   virtual bool Predict(cv::Mat* im, MattingResult* result);
 

diff --git a/examples/vision/matting/modnet/python/README.md b/examples/vision/matting/modnet/python/README.md
@@ -33,6 +33,7 @@ python infer.py --model modnet_photographic_portrait_matting.onnx --image mattin
 <img width="200" height="200" float="left" src="https://user-images.githubusercontent.com/67993288/186852116-cf91445b-3a67-45d9-a675-c69fe77c383a.jpg">
 <img width="200" height="200" float="left" src="https://user-images.githubusercontent.com/67993288/186851964-4c9086b9-3490-4fcb-82f9-2106c63aa4f3.jpg">
 </div>
+
 ## MODNet Python接口
 
 ```python

diff --git a/examples/vision/matting/ppmatting/README.md b/examples/vision/matting/ppmatting/README.md
@@ -15,19 +15,18 @@
 
 在部署前，需要先将PPMatting导出成部署模型，导出步骤参考文档[导出模型](https://github.com/PaddlePaddle/PaddleSeg/tree/release/2.6/Matting)
 
-注意：在导出模型时不要进行NMS的去除操作，正常导出即可。
 
 ## 下载预训练模型
 
 为了方便开发者的测试，下面提供了PPMatting导出的各系列模型，开发者可直接下载使用。
 
-其中精度指标来源于PPMatting中对各模型的介绍，详情各参考PPMatting中的说明。
+其中精度指标来源于PPMatting中对各模型的介绍(未提供精度数据)，详情各参考PPMatting中的说明。
 
 
 | 模型                                                               | 参数大小    | 精度    | 备注 |
 |:---------------------------------------------------------------- |:----- |:----- | :------ |
-| [PPMatting](https://bj.bcebos.com/paddlehub/fastdeploy/PP-Matting-512.tgz) | 87MB | - |
-| [PPMatting](https://bj.bcebos.com/paddlehub/fastdeploy/PP-Matting-1024.tgz) | 87MB | - |
+| [PPMatting-512](https://bj.bcebos.com/paddlehub/fastdeploy/PP-Matting-512.tgz) | 87MB | - |
+| [PPMatting-1024](https://bj.bcebos.com/paddlehub/fastdeploy/PP-Matting-1024.tgz) | 87MB | - |
 
 
 

diff --git a/examples/vision/matting/ppmatting/cpp/README.md b/examples/vision/matting/ppmatting/cpp/README.md
@@ -27,7 +27,7 @@ wget https://bj.bcebos.com/paddlehub/fastdeploy/matting_bgr.jpg
 
 # CPU推理
 ./infer_demo PP-Matting-512 matting_input.jpg matting_bgr.jpg 0
-# GPU推理 (TODO: ORT-GPU 推理会报错)
+# GPU推理
 ./infer_demo PP-Matting-512 matting_input.jpg matting_bgr.jpg 1
 # GPU上TensorRT推理
 ./infer_demo PP-Matting-512 matting_input.jpg matting_bgr.jpg 2

diff --git a/examples/vision/matting/ppmatting/cpp/infer.cc b/examples/vision/matting/ppmatting/cpp/infer.cc
@@ -25,8 +25,9 @@ void CpuInfer(const std::string& model_dir, const std::string& image_file,
   auto model_file = model_dir + sep + "model.pdmodel";
   auto params_file = model_dir + sep + "model.pdiparams";
   auto config_file = model_dir + sep + "deploy.yaml";
+  auto option = fastdeploy::RuntimeOption();
   auto model = fastdeploy::vision::matting::PPMatting(model_file, params_file,
-                                                      config_file);
+                                                      config_file, option);
   if (!model.Initialized()) {
     std::cerr << "Failed to initialize." << std::endl;
     return;
@@ -58,6 +59,7 @@ void GpuInfer(const std::string& model_dir, const std::string& image_file,
 
   auto option = fastdeploy::RuntimeOption();
   option.UseGpu();
+  option.UsePaddleBackend();
   auto model = fastdeploy::vision::matting::PPMatting(model_file, params_file,
                                                       config_file, option);
   if (!model.Initialized()) {

diff --git a/examples/vision/matting/ppmatting/python/README.md b/examples/vision/matting/ppmatting/python/README.md
@@ -19,7 +19,7 @@ wget https://bj.bcebos.com/paddlehub/fastdeploy/matting_input.jpg
 wget https://bj.bcebos.com/paddlehub/fastdeploy/matting_bgr.jpg
 # CPU推理
 python infer.py --model PP-Matting-512 --image matting_input.jpg --bg matting_bgr.jpg --device cpu
-# GPU推理 (TODO: ORT-GPU 推理会报错)
+# GPU推理
 python infer.py --model PP-Matting-512 --image matting_input.jpg --bg matting_bgr.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 infer.py --model PP-Matting-512 --image matting_input.jpg --bg matting_bgr.jpg --device gpu --use_trt True

diff --git a/examples/vision/matting/ppmatting/python/infer.py b/examples/vision/matting/ppmatting/python/infer.py
@@ -32,13 +32,14 @@ def parse_arguments():
 
 def build_option(args):
     option = fd.RuntimeOption()
-
     if args.device.lower() == "gpu":
         option.use_gpu()
+        option.use_paddle_backend()
 
     if args.use_trt:
         option.use_trt_backend()
         option.set_trt_input_shape("img", [1, 3, 512, 512])
+
     return option
 
 

diff --git a/external/paddle_inference.cmake b/external/paddle_inference.cmake
@@ -46,6 +46,7 @@ else()
   set(OMP_LIB "${PADDLEINFERENCE_INSTALL_DIR}/third_party/install/mklml/lib/libiomp5.so")
 endif(WIN32)
 
+
 set(PADDLEINFERENCE_URL_BASE "https://bj.bcebos.com/fastdeploy/third_libs/")
 set(PADDLEINFERENCE_VERSION "2.4-dev")
 if(WIN32)