Add new model PaddleSeg (#30)

* Support new model PaddleSeg

* Fix conflict

* PaddleSeg add visulization function

* fix bug

* Fix BindPPSeg wrong name

* Fix variable name

* Update by comments

* Add ppseg-unet example python version

Co-authored-by: Jason <jiangjiajun@baidu.com>

examples/vision/ppseg_unet.cc

@@ -0,0 +1,59 @@
+// 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"
+#include "yaml-cpp/yaml.h"
+
+int main() {
+  namespace vis = fastdeploy::vision;
+
+  std::string model_file = "../resources/models/unet_Cityscapes/model.pdmodel";
+  std::string params_file =
+      "../resources/models/unet_Cityscapes/model.pdiparams";
+  std::string config_file = "../resources/models/unet_Cityscapes/deploy.yaml";
+  std::string img_path = "../resources/images/cityscapes_demo.png";
+  std::string vis_path = "../resources/outputs/vis.jpeg";
+
+  auto model = vis::ppseg::Model(model_file, params_file, config_file);
+  if (!model.Initialized()) {
+    std::cerr << "Init Failed." << std::endl;
+    return -1;
+  }
+
+  cv::Mat im = cv::imread(img_path);
+  cv::Mat vis_im;
+
+  vis::SegmentationResult res;
+  if (!model.Predict(&im, &res)) {
+    std::cerr << "Prediction Failed." << std::endl;
+    return -1;
+  } else {
+    std::cout << "Prediction Done!" << std::endl;
+  }
+
+  // 输出预测框结果
+  std::cout << res.Str() << std::endl;
+
+  YAML::Node cfg = YAML::LoadFile(config_file);
+  int num_classes = 19;
+  if (cfg["Deploy"]["num_classes"]) {
+    num_classes = cfg["Deploy"]["num_classes"].as<int>();
+  }
+
+  // 可视化预测结果
+  vis::Visualize::VisSegmentation(im, res, &vis_im, num_classes);
+  cv::imwrite(vis_path, vis_im);
+  std::cout << "Inference Done! Saved: " << vis_path << std::endl;
+  return 0;
+}

fastdeploy/vision.h

@@ -19,6 +19,7 @@
 #include "fastdeploy/vision/meituan/yolov6.h"
 #include "fastdeploy/vision/ppcls/model.h"
 #include "fastdeploy/vision/ppdet/ppyoloe.h"
+#include "fastdeploy/vision/ppseg/model.h"
 #include "fastdeploy/vision/ultralytics/yolov5.h"
 #include "fastdeploy/vision/wongkinyiu/yolor.h"
 #include "fastdeploy/vision/wongkinyiu/yolov7.h"

fastdeploy/vision/__init__.py

@@ -16,6 +16,7 @@
 from . import evaluation
 from . import ppcls
 from . import ppdet
+from . import ppseg
 from . import ultralytics
 from . import meituan
 from . import megvii

fastdeploy/vision/common/result.cc

@@ -72,5 +72,27 @@ std::string DetectionResult::Str() {
   return out;
 }
 
+void SegmentationResult::Clear() {
+  std::vector<std::vector<int64_t>>().swap(masks);
+}
+
+void SegmentationResult::Resize(int64_t height, int64_t width) {
+  masks.resize(height, std::vector<int64_t>(width));
+}
+
+std::string SegmentationResult::Str() {
+  std::string out;
+  out = "SegmentationResult Image masks 10 rows x 10 cols: \n";
+  for (size_t i = 0; i < 10; ++i) {
+    out += "[";
+    for (size_t j = 0; j < 10; ++j) {
+      out = out + std::to_string(masks[i][j]) + ", ";
+    }
+    out += ".....]\n";
+  }
+  out += "...........\n";
+  return out;
+}
+
 }  // namespace vision
 }  // namespace fastdeploy

fastdeploy/vision/common/result.h

@@ -56,5 +56,18 @@ struct FASTDEPLOY_DECL DetectionResult : public BaseResult {
   std::string Str();
 };
 
+struct FASTDEPLOY_DECL SegmentationResult : public BaseResult {
+  // mask
+  std::vector<std::vector<int64_t>> masks;
+
+  ResultType type = ResultType::SEGMENTATION;
+
+  void Clear();
+
+  void Resize(int64_t height, int64_t width);
+
+  std::string Str();
+};
+
 }  // namespace vision
 }  // namespace fastdeploy

fastdeploy/vision/ppseg/__init__.py

@@ -0,0 +1,37 @@
+# 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.
+
+from __future__ import absolute_import
+import logging
+from ... import FastDeployModel, Frontend
+from ... import fastdeploy_main as C
+
+
+class Model(FastDeployModel):
+    def __init__(self,
+                 model_file,
+                 params_file,
+                 config_file,
+                 backend_option=None,
+                 model_format=Frontend.PADDLE):
+        super(Model, self).__init__(backend_option)
+
+        assert model_format == Frontend.PADDLE, "PaddleSeg only support model format of Frontend.Paddle now."
+        self._model = C.vision.ppseg.Model(model_file, params_file,
+                                           config_file, self._runtime_option,
+                                           model_format)
+        assert self.initialized, "PaddleSeg model initialize failed."
+
+    def predict(self, input_image):
+        return self._model.predict(input_image)

fastdeploy/vision/ppseg/model.cc

@@ -0,0 +1,140 @@
+#include "fastdeploy/vision/ppseg/model.h"
+#include "fastdeploy/vision.h"
+#include "fastdeploy/vision/utils/utils.h"
+#include "yaml-cpp/yaml.h"
+
+namespace fastdeploy {
+namespace vision {
+namespace ppseg {
+
+Model::Model(const std::string& model_file, const std::string& params_file,
+             const std::string& config_file, const RuntimeOption& custom_option,
+             const Frontend& model_format) {
+  config_file_ = config_file;
+  valid_cpu_backends = {Backend::ORT, Backend::PDINFER};
+  valid_gpu_backends = {Backend::ORT, Backend::PDINFER};
+  runtime_option = custom_option;
+  runtime_option.model_format = model_format;
+  runtime_option.model_file = model_file;
+  runtime_option.params_file = params_file;
+  initialized = Initialize();
+}
+
+bool Model::Initialize() {
+  if (!BuildPreprocessPipelineFromConfig()) {
+    FDERROR << "Failed to build preprocess pipeline from configuration file."
+            << std::endl;
+    return false;
+  }
+  if (!InitRuntime()) {
+    FDERROR << "Failed to initialize fastdeploy backend." << std::endl;
+    return false;
+  }
+  return true;
+}
+
+bool Model::BuildPreprocessPipelineFromConfig() {
+  processors_.clear();
+  YAML::Node cfg;
+  processors_.push_back(std::make_shared<BGR2RGB>());
+  try {
+    cfg = YAML::LoadFile(config_file_);
+  } catch (YAML::BadFile& e) {
+    FDERROR << "Failed to load yaml file " << config_file_
+            << ", maybe you should check this file." << std::endl;
+    return false;
+  }
+
+  if (cfg["Deploy"]["transforms"]) {
+    auto preprocess_cfg = cfg["Deploy"]["transforms"];
+    for (const auto& op : preprocess_cfg) {
+      FDASSERT(op.IsMap(),
+               "Require the transform information in yaml be Map type.");
+      if (op["type"].as<std::string>() == "Normalize") {
+        std::vector<float> mean = {0.5, 0.5, 0.5};
+        std::vector<float> std = {0.5, 0.5, 0.5};
+        if (op["mean"]) {
+          mean = op["mean"].as<std::vector<float>>();
+        }
+        if (op["std"]) {
+          std = op["std"].as<std::vector<float>>();
+        }
+        processors_.push_back(std::make_shared<Normalize>(mean, std));
+
+      } else if (op["type"].as<std::string>() == "Resize") {
+        const auto& target_size = op["target_size"];
+        int resize_width = target_size[0].as<int>();
+        int resize_height = target_size[1].as<int>();
+        processors_.push_back(
+            std::make_shared<Resize>(resize_width, resize_height));
+      }
+    }
+    processors_.push_back(std::make_shared<HWC2CHW>());
+  }
+  return true;
+}
+
+bool Model::Preprocess(Mat* mat, FDTensor* output) {
+  for (size_t i = 0; i < processors_.size(); ++i) {
+    if (!(*(processors_[i].get()))(mat)) {
+      FDERROR << "Failed to process image data in " << processors_[i]->Name()
+              << "." << std::endl;
+      return false;
+    }
+  }
+  mat->ShareWithTensor(output);
+  output->shape.insert(output->shape.begin(), 1);
+  output->name = InputInfoOfRuntime(0).name;
+  return true;
+}
+
+bool Model::Postprocess(const FDTensor& infer_result,
+                        SegmentationResult* result) {
+  FDASSERT(infer_result.dtype == FDDataType::INT64,
+           "Require the data type of output is int64, but now it's " +
+               Str(const_cast<fastdeploy::FDDataType&>(infer_result.dtype)) +
+               ".");
+  result->Clear();
+  std::vector<int64_t> output_shape = infer_result.shape;
+  int out_num = std::accumulate(output_shape.begin(), output_shape.end(), 1,
+                                std::multiplies<int>());
+  const int64_t* infer_result_buffer =
+      reinterpret_cast<const int64_t*>(infer_result.data.data());
+  int64_t height = output_shape[1];
+  int64_t width = output_shape[2];
+  result->Resize(height, width);
+  for (int64_t i = 0; i < height; i++) {
+    int64_t begin = i * width;
+    int64_t end = (i + 1) * width - 1;
+    std::copy(infer_result_buffer + begin, infer_result_buffer + end,
+              result->masks[i].begin());
+  }
+
+  return true;
+}
+
+bool Model::Predict(cv::Mat* im, SegmentationResult* result) {
+  Mat mat(*im);
+  std::vector<FDTensor> processed_data(1);
+  if (!Preprocess(&mat, &(processed_data[0]))) {
+    FDERROR << "Failed to preprocess input data while using model:"
+            << ModelName() << "." << std::endl;
+    return false;
+  }
+  std::vector<FDTensor> infer_result(1);
+  if (!Infer(processed_data, &infer_result)) {
+    FDERROR << "Failed to inference while using model:" << ModelName() << "."
+            << std::endl;
+    return false;
+  }
+  if (!Postprocess(infer_result[0], result)) {
+    FDERROR << "Failed to postprocess while using model:" << ModelName() << "."
+            << std::endl;
+    return false;
+  }
+  return true;
+}
+
+}  // namespace ppseg
+}  // namespace vision
+}  // namespace fastdeploy

fastdeploy/vision/ppseg/model.h

@@ -0,0 +1,35 @@
+#pragma once
+#include "fastdeploy/fastdeploy_model.h"
+#include "fastdeploy/vision/common/processors/transform.h"
+#include "fastdeploy/vision/common/result.h"
+
+namespace fastdeploy {
+namespace vision {
+namespace ppseg {
+
+class FASTDEPLOY_DECL Model : public FastDeployModel {
+ public:
+  Model(const std::string& model_file, const std::string& params_file,
+        const std::string& config_file,
+        const RuntimeOption& custom_option = RuntimeOption(),
+        const Frontend& model_format = Frontend::PADDLE);
+
+  std::string ModelName() const { return "ppseg"; }
+
+  virtual bool Predict(cv::Mat* im, SegmentationResult* result);
+
+ private:
+  bool Initialize();
+
+  bool BuildPreprocessPipelineFromConfig();
+
+  bool Preprocess(Mat* mat, FDTensor* outputs);
+
+  bool Postprocess(const FDTensor& infer_result, SegmentationResult* result);
+
+  std::vector<std::shared_ptr<Processor>> processors_;
+  std::string config_file_;
+};
+}  // namespace ppseg
+}  // namespace vision
+}  // namespace fastdeploy

fastdeploy/vision/ppseg/ppseg_pybind.cc

@@ -0,0 +1,30 @@
+// 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/pybind/main.h"
+
+namespace fastdeploy {
+void BindPPSeg(pybind11::module& m) {
+  auto ppseg_module =
+      m.def_submodule("ppseg", "Module to deploy PaddleSegmentation.");
+  pybind11::class_<vision::ppseg::Model, FastDeployModel>(ppseg_module, "Model")
+      .def(pybind11::init<std::string, std::string, std::string, RuntimeOption,
+                          Frontend>())
+      .def("predict", [](vision::ppseg::Model& self, pybind11::array& data) {
+        auto mat = PyArrayToCvMat(data);
+        vision::SegmentationResult res;
+        self.Predict(&mat, &res);
+        return res;
+      });
+}
+}  // namespace fastdeploy

fastdeploy/vision/vision_pybind.cc

@@ -19,6 +19,7 @@ namespace fastdeploy {
 void BindPPCls(pybind11::module& m);
 void BindPPDet(pybind11::module& m);
 void BindWongkinyiu(pybind11::module& m);
+void BindPPSeg(pybind11::module& m);
 void BindUltralytics(pybind11::module& m);
 void BindMeituan(pybind11::module& m);
 void BindMegvii(pybind11::module& m);
@@ -42,8 +43,15 @@ void BindVision(pybind11::module& m) {
       .def("__repr__", &vision::DetectionResult::Str)
       .def("__str__", &vision::DetectionResult::Str);
 
+  pybind11::class_<vision::SegmentationResult>(m, "SegmentationResult")
+      .def(pybind11::init())
+      .def_readwrite("masks", &vision::SegmentationResult::masks)
+      .def("__repr__", &vision::SegmentationResult::Str)
+      .def("__str__", &vision::SegmentationResult::Str);
+
   BindPPCls(m);
   BindPPDet(m);
+  BindPPSeg(m);
   BindUltralytics(m);
   BindWongkinyiu(m);
   BindMeituan(m);