[Model] Refactor PaddleClas module (#505)

* Refactor the PaddleClas module

* fix bug

* remove debug code

* clean unused code

* support pybind

* Update fd_tensor.h

* Update fd_tensor.cc

* temporary revert python api

* fix ci error

* fix code style problem

fastdeploy/vision/classification/ppcls/model.cc

@@ -14,9 +14,6 @@
 
 #include "fastdeploy/vision/classification/ppcls/model.h"
 
-#include "fastdeploy/vision/utils/utils.h"
-#include "yaml-cpp/yaml.h"
-
 namespace fastdeploy {
 namespace vision {
 namespace classification {
@@ -25,8 +22,7 @@ PaddleClasModel::PaddleClasModel(const std::string& model_file,
                                  const std::string& params_file,
                                  const std::string& config_file,
                                  const RuntimeOption& custom_option,
-                                 const ModelFormat& model_format) {
+                                 const ModelFormat& model_format) : preprocessor_(config_file) {
-  config_file_ = config_file;
   valid_cpu_backends = {Backend::ORT, Backend::OPENVINO, Backend::PDINFER,
                         Backend::LITE};
   valid_gpu_backends = {Backend::ORT, Backend::PDINFER, Backend::TRT};
@@ -38,11 +34,6 @@ PaddleClasModel::PaddleClasModel(const std::string& model_file,
 }
 
 bool PaddleClasModel::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;
@@ -50,105 +41,41 @@ bool PaddleClasModel::Initialize() {
   return true;
 }
 
-bool PaddleClasModel::BuildPreprocessPipelineFromConfig() {
-  processors_.clear();
-  YAML::Node cfg;
-  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;
-  }
-  auto preprocess_cfg = cfg["PreProcess"]["transform_ops"];
-  processors_.push_back(std::make_shared<BGR2RGB>());
-  for (const auto& op : preprocess_cfg) {
-    FDASSERT(op.IsMap(),
-             "Require the transform information in yaml be Map type.");
-    auto op_name = op.begin()->first.as<std::string>();
-    if (op_name == "ResizeImage") {
-      int target_size = op.begin()->second["resize_short"].as<int>();
-      bool use_scale = false;
-      int interp = 1;
-      processors_.push_back(
-          std::make_shared<ResizeByShort>(target_size, 1, use_scale));
-    } else if (op_name == "CropImage") {
-      int width = op.begin()->second["size"].as<int>();
-      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));
-    } else if (op_name == "ToCHWImage") {
-      processors_.push_back(std::make_shared<HWC2CHW>());
-    } else {
-      FDERROR << "Unexcepted preprocess operator: " << op_name << "."
-              << std::endl;
-      return false;
-    }
-  }
-  return true;
-}
-
-bool PaddleClasModel::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;
-    }
-  }
-
-  int channel = mat->Channels();
-  int width = mat->Width();
-  int height = mat->Height();
-  output->name = InputInfoOfRuntime(0).name;
-  output->SetExternalData({1, channel, height, width}, FDDataType::FP32,
-                          mat->Data());
-  return true;
-}
-
-bool PaddleClasModel::Postprocess(const FDTensor& infer_result,
-                                  ClassifyResult* result, int topk) {
-  int num_classes = infer_result.shape[1];
-  const float* infer_result_buffer =
-      reinterpret_cast<const float*>(infer_result.Data());
-  topk = std::min(num_classes, topk);
-  result->label_ids =
-      utils::TopKIndices(infer_result_buffer, num_classes, topk);
-  result->scores.resize(topk);
-  for (int i = 0; i < topk; ++i) {
-    result->scores[i] = *(infer_result_buffer + result->label_ids[i]);
-  }
-  return true;
-}
-
 bool PaddleClasModel::Predict(cv::Mat* im, ClassifyResult* result, int topk) {
-  Mat mat(*im);
+  postprocessor_.SetTopk(topk);
-  std::vector<FDTensor> processed_data(1);
+  if (!Predict(*im, result)) {
-  if (!Preprocess(&mat, &(processed_data[0]))) {
+    return false;
-    FDERROR << "Failed to preprocess input data while using model:"
+  }
-            << ModelName() << "." << std::endl;
+  return true;
+}
+
+bool PaddleClasModel::Predict(const cv::Mat& im, ClassifyResult* result) {
+  std::vector<ClassifyResult> results;
+  if (!BatchPredict({im}, &results)) {
+    return false;
+  }
+  *result = std::move(results[0]);
+  return true;
+}
+
+bool PaddleClasModel::BatchPredict(const std::vector<cv::Mat>& images, std::vector<ClassifyResult>* results) {
+  std::vector<FDMat> fd_images = WrapMat(images);
+  if (!preprocessor_.Run(&fd_images, &reused_input_tensors)) {
+    FDERROR << "Failed to preprocess the input image." << std::endl;
     return false;
   }
 
-  std::vector<FDTensor> infer_result(1);
+  reused_input_tensors[0].name = InputInfoOfRuntime(0).name;
-  if (!Infer(processed_data, &infer_result)) {
+  if (!Infer(reused_input_tensors, &reused_output_tensors)) {
-    FDERROR << "Failed to inference while using model:" << ModelName() << "."
+    FDERROR << "Failed to inference by runtime." << std::endl;
-            << std::endl;
     return false;
   }
 
-  if (!Postprocess(infer_result[0], result, topk)) {
+  if (!postprocessor_.Run(reused_output_tensors, results)) {
-    FDERROR << "Failed to postprocess while using model:" << ModelName() << "."
+    FDERROR << "Failed to postprocess the inference results by runtime." << std::endl;
-            << std::endl;
     return false;
   }
+
   return true;
 }
 

fastdeploy/vision/classification/ppcls/model.h

@@ -14,8 +14,8 @@
 
 #pragma once
 #include "fastdeploy/fastdeploy_model.h"
-#include "fastdeploy/vision/common/processors/transform.h"
+#include "fastdeploy/vision/classification/ppcls/preprocessor.h"
-#include "fastdeploy/vision/common/result.h"
+#include "fastdeploy/vision/classification/ppcls/postprocessor.h"
 
 namespace fastdeploy {
 namespace vision {
@@ -43,28 +43,46 @@ class FASTDEPLOY_DECL PaddleClasModel : public FastDeployModel {
   /// Get model's name
   virtual std::string ModelName() const { return "PaddleClas/Model"; }
 
-  /** \brief Predict the classification result for an input image
+  /** \brief DEPRECATED Predict the classification result for an input image, remove at 1.0 version
    *
    * \param[in] im The input image data, comes from cv::imread()
    * \param[in] result The output classification result will be writen to this structure
-   * \param[in] topk (int)The topk result by the classify confidence score, default 1
    * \return true if the prediction successed, otherwise false
    */
-  // TODO(jiangjiajun) Batch is on the way
   virtual bool Predict(cv::Mat* im, ClassifyResult* result, int topk = 1);
 
+
+  /** \brief Predict the classification result for an input image
+   *
+   * \param[in] img The input image data, comes from cv::imread()
+   * \param[in] result The output classification result
+   * \return true if the prediction successed, otherwise false
+   */
+  virtual bool Predict(const cv::Mat& img, ClassifyResult* result);
+
+  /** \brief Predict the classification results for a batch of input images
+   *
+   * \param[in] imgs, The input image list, each element comes from cv::imread()
+   * \param[in] results The output classification result list
+   * \return true if the prediction successed, otherwise false
+   */
+  virtual bool BatchPredict(const std::vector<cv::Mat>& imgs,
+                            std::vector<ClassifyResult>* results);
+
+  /// Get preprocessor reference of PaddleClasModel
+  virtual PaddleClasPreprocessor& GetPreprocessor() {
+    return preprocessor_;
+  }
+
+  /// Get postprocessor reference of PaddleClasModel
+  virtual PaddleClasPostprocessor& GetPostprocessor() {
+    return postprocessor_;
+  }
+
  protected:
   bool Initialize();
-
+  PaddleClasPreprocessor preprocessor_;
-  bool BuildPreprocessPipelineFromConfig();
+  PaddleClasPostprocessor postprocessor_;
-
-  bool Preprocess(Mat* mat, FDTensor* outputs);
-
-  bool Postprocess(const FDTensor& infer_result, ClassifyResult* result,
-                   int topk = 1);
-
-  std::vector<std::shared_ptr<Processor>> processors_;
-  std::string config_file_;
 };
 
 typedef PaddleClasModel PPLCNet;

fastdeploy/vision/classification/ppcls/postprocessor.cc

@@ -0,0 +1,53 @@
+// 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/classification/ppcls/postprocessor.h"
+#include "fastdeploy/vision/utils/utils.h"
+
+namespace fastdeploy {
+namespace vision {
+namespace classification {
+
+PaddleClasPostprocessor::PaddleClasPostprocessor(int topk) {
+  topk_ = topk;
+  initialized_ = true;
+}
+
+bool PaddleClasPostprocessor::Run(const std::vector<FDTensor>& infer_result, std::vector<ClassifyResult>* results) {
+  if (!initialized_) {
+    FDERROR << "Postprocessor is not initialized." << std::endl;
+    return false;
+  }
+
+  int batch = infer_result[0].shape[0];
+  int num_classes = infer_result[0].shape[1];
+  const float* infer_result_data = reinterpret_cast<const float*>(infer_result[0].Data());
+ 
+  results->resize(batch);
+
+  int topk = std::min(num_classes, topk_);
+  for (int i = 0; i < batch; ++i) {
+    (*results)[i].label_ids = utils::TopKIndices(infer_result_data + i * num_classes, num_classes, topk);
+    (*results)[i].scores.resize(topk);
+    for (int j = 0; j < topk; ++j) {
+      (*results)[i].scores[j] = infer_result_data[i * num_classes + (*results)[i].label_ids[j]];
+    }
+  }
+
+  return true;
+}
+
+}  // namespace classification
+}  // namespace vision
+}  // namespace fastdeploy

fastdeploy/vision/classification/ppcls/postprocessor.h

@@ -0,0 +1,55 @@
+// 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/transform.h"
+#include "fastdeploy/vision/common/result.h"
+
+namespace fastdeploy {
+namespace vision {
+
+namespace classification {
+/*! @brief Postprocessor object for PaddleClas serials model.
+ */
+class FASTDEPLOY_DECL PaddleClasPostprocessor {
+ public:
+  /** \brief Create a postprocessor instance for PaddleClas serials model
+   *
+   * \param[in] topk The topk result filtered by the classify confidence score, default 1
+   */
+  explicit PaddleClasPostprocessor(int topk = 1);
+
+  /** \brief Process the result of runtime and fill to ClassifyResult structure
+   *
+   * \param[in] tensors The inference result from runtime
+   * \param[in] result The output result of classification
+   * \return true if the postprocess successed, otherwise false
+   */
+  bool Run(const std::vector<FDTensor>& tensors,
+          std::vector<ClassifyResult>* result);
+
+  /// Set topk value
+  void SetTopk(int topk) { topk_ = topk; }
+
+  /// Get topk value
+  int GetTopk() const { return topk_; }
+
+ private:
+  int topk_ = 1;
+  bool initialized_ = false;
+};
+
+}  // namespace classification
+}  // namespace vision
+}  // namespace fastdeploy

fastdeploy/vision/classification/ppcls/ppcls_pybind.cc

@@ -15,16 +15,62 @@
 
 namespace fastdeploy {
 void BindPaddleClas(pybind11::module& m) {
+  pybind11::class_<vision::classification::PaddleClasPreprocessor>(
+      m, "PaddleClasPreprocessor")
+      .def(pybind11::init<std::string>())
+      .def("run", [](vision::classification::PaddleClasPreprocessor& self, std::vector<pybind11::array>& im_list) {
+        std::vector<vision::FDMat> images;
+        for (size_t i = 0; i < im_list.size(); ++i) {
+          images.push_back(vision::WrapMat(PyArrayToCvMat(im_list[i])));
+        }
+        std::vector<FDTensor> outputs;
+        if (!self.Run(&images, &outputs)) {
+          pybind11::eval("raise Exception('Failed to preprocess the input data in PaddleClasPreprocessor.')");
+        }
+        return outputs;
+      });
+
+  pybind11::class_<vision::classification::PaddleClasPostprocessor>(
+      m, "PaddleClasPostprocessor")
+      .def(pybind11::init<int>())
+      .def("run", [](vision::classification::PaddleClasPostprocessor& self, std::vector<FDTensor>& inputs) {
+        std::vector<vision::ClassifyResult> results;
+        if (!self.Run(inputs, &results)) {
+          pybind11::eval("raise Exception('Failed to postprocess the runtime result in PaddleClasPostprocessor.')");
+        }
+        return results;
+      })
+      .def("run", [](vision::classification::PaddleClasPostprocessor& self, std::vector<pybind11::array>& input_array) {
+        std::vector<vision::ClassifyResult> results;
+        std::vector<FDTensor> inputs;
+        PyArrayToTensorList(input_array, &inputs, /*share_buffer=*/true);
+        if (!self.Run(inputs, &results)) {
+          pybind11::eval("raise Exception('Failed to postprocess the runtime result in PaddleClasPostprocessor.')");
+        }
+        return results;
+      })
+      .def_property("topk", &vision::classification::PaddleClasPostprocessor::GetTopk, &vision::classification::PaddleClasPostprocessor::SetTopk);
+
   pybind11::class_<vision::classification::PaddleClasModel, FastDeployModel>(
       m, "PaddleClasModel")
       .def(pybind11::init<std::string, std::string, std::string, RuntimeOption,
                           ModelFormat>())
-      .def("predict", [](vision::classification::PaddleClasModel& self,
+      .def("predict", [](vision::classification::PaddleClasModel& self, pybind11::array& data) {
-                         pybind11::array& data, int topk = 1) {
+        cv::Mat im = PyArrayToCvMat(data);
-        auto mat = PyArrayToCvMat(data);
+        vision::ClassifyResult result;
-        vision::ClassifyResult res;
+        self.Predict(im, &result);
-        self.Predict(&mat, &res, topk);
+        return result;
-        return res;
+      })
-      });
+      .def("batch_predict", [](vision::classification::PaddleClasModel& self, std::vector<pybind11::array>& data) {
+        std::vector<cv::Mat> images;
+        for (size_t i = 0; i < data.size(); ++i) {
+          images.push_back(PyArrayToCvMat(data[i]));
+        }
+        std::vector<vision::ClassifyResult> results;
+        self.BatchPredict(images, &results);
+        return results;
+      })
+      .def_property_readonly("preprocessor", &vision::classification::PaddleClasModel::GetPreprocessor)
+      .def_property_readonly("postprocessor", &vision::classification::PaddleClasModel::GetPostprocessor);
 }
 }  // namespace fastdeploy

fastdeploy/vision/classification/ppcls/preprocessor.cc

@@ -0,0 +1,108 @@
+// 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/classification/ppcls/preprocessor.h"
+#include "fastdeploy/function/concat.h"
+#include "yaml-cpp/yaml.h"
+
+namespace fastdeploy {
+namespace vision {
+namespace classification {
+
+PaddleClasPreprocessor::PaddleClasPreprocessor(const std::string& config_file) {
+  FDASSERT(BuildPreprocessPipelineFromConfig(config_file), "Failed to create PaddleClasPreprocessor.");
+  initialized_ = true;
+}
+
+bool PaddleClasPreprocessor::BuildPreprocessPipelineFromConfig(const std::string& config_file) {
+  processors_.clear();
+  YAML::Node cfg;
+  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;
+  }
+  auto preprocess_cfg = cfg["PreProcess"]["transform_ops"];
+  processors_.push_back(std::make_shared<BGR2RGB>());
+  for (const auto& op : preprocess_cfg) {
+    FDASSERT(op.IsMap(),
+             "Require the transform information in yaml be Map type.");
+    auto op_name = op.begin()->first.as<std::string>();
+    if (op_name == "ResizeImage") {
+      int target_size = op.begin()->second["resize_short"].as<int>();
+      bool use_scale = false;
+      int interp = 1;
+      processors_.push_back(
+          std::make_shared<ResizeByShort>(target_size, 1, use_scale));
+    } else if (op_name == "CropImage") {
+      int width = op.begin()->second["size"].as<int>();
+      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));
+    } else if (op_name == "ToCHWImage") {
+      processors_.push_back(std::make_shared<HWC2CHW>());
+    } else {
+      FDERROR << "Unexcepted preprocess operator: " << op_name << "."
+              << std::endl;
+      return false;
+    }
+  }
+
+  // Fusion will improve performance
+  FuseTransforms(&processors_);
+  return true;
+}
+
+bool PaddleClasPreprocessor::Run(std::vector<FDMat>* images, std::vector<FDTensor>* outputs) {
+  if (!initialized_) {
+    FDERROR << "The preprocessor is not initialized." << std::endl;
+    return false;
+  }
+  if (images->size() == 0) {
+    FDERROR << "The size of input images should be greater than 0." << std::endl;
+    return false;
+  }
+
+  for (size_t i = 0; i < images->size(); ++i) {
+    for (size_t j = 0; j < processors_.size(); ++j) {
+      if (!(*(processors_[j].get()))(&((*images)[i]))) {
+        FDERROR << "Failed to processs image:" << i << " in " << processors_[i]->Name() << "." << std::endl;
+        return false;
+      }
+    }
+  }
+
+  outputs->resize(1);
+  // Concat all the preprocessed data to a batch tensor
+  std::vector<FDTensor> tensors(images->size()); 
+  for (size_t i = 0; i < images->size(); ++i) {
+    (*images)[i].ShareWithTensor(&(tensors[i]));
+    tensors[i].ExpandDim(0);
+  }
+  Concat(tensors, &((*outputs)[0]), 0);
+  return true;
+}
+
+}  // namespace classification
+}  // namespace vision
+}  // namespace fastdeploy

fastdeploy/vision/classification/ppcls/preprocessor.h

@@ -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.
+
+#pragma once
+#include "fastdeploy/vision/common/processors/transform.h"
+#include "fastdeploy/vision/common/result.h"
+
+namespace fastdeploy {
+namespace vision {
+
+namespace classification {
+/*! @brief Preprocessor object for PaddleClas serials model.
+ */
+class FASTDEPLOY_DECL PaddleClasPreprocessor {
+ public:
+  /** \brief Create a preprocessor instance for PaddleClas serials model
+   *
+   * \param[in] config_file Path of configuration file for deployment, e.g resnet/infer_cfg.yml
+   */
+  explicit PaddleClasPreprocessor(const std::string& config_file);
+
+  /** \brief Process the input image and prepare input tensors for runtime
+   *
+   * \param[in] images The input image data list, all the elements are returned by cv::imread()
+   * \param[in] outputs The output tensors which will feed in runtime
+   * \return true if the preprocess successed, otherwise false
+   */
+  bool Run(std::vector<FDMat>* images, std::vector<FDTensor>* outputs);
+
+
+ private:
+  bool BuildPreprocessPipelineFromConfig(const std::string& config_file);
+  std::vector<std::shared_ptr<Processor>> processors_;
+  bool initialized_ = false;
+};
+
+}  // namespace classification
+}  // namespace vision
+}  // namespace fastdeploy

fastdeploy/vision/common/processors/resize_by_short.cc

@@ -51,7 +51,7 @@ bool ResizeByShort::ImplByFlyCV(Mat* mat) {
   } else if (interp_ == 2) {
     interp_method = fcv::InterpolationType::INTER_CUBIC;
   } else {
-    FDERROR << "LimitLong: Only support interp_ be 0/1/2 with FlyCV, but "
+    FDERROR << "LimitByShort: Only support interp_ be 0/1/2 with FlyCV, but "
                "now it's "
             << interp_ << "." << std::endl;
     return false;

fastdeploy/vision/common/result.cc

@@ -35,6 +35,14 @@ std::string ClassifyResult::Str() {
   return out;
 }
 
+ClassifyResult& ClassifyResult::operator=(ClassifyResult&& other) {
+  if (&other != this) {
+    label_ids = std::move(other.label_ids);
+    scores = std::move(other.scores);
+  }
+  return *this;
+}
+
 void Mask::Reserve(int size) { data.reserve(size); }
 
 void Mask::Resize(int size) { data.resize(size); }

fastdeploy/vision/common/result.h

@@ -44,6 +44,7 @@ struct FASTDEPLOY_DECL BaseResult {
 /*! @brief Classify result structure for all the image classify models
  */
 struct FASTDEPLOY_DECL ClassifyResult : public BaseResult {
+  ClassifyResult() = default;
   /// Classify result for an image
   std::vector<int32_t> label_ids;
   /// The confidence for each classify result
@@ -53,6 +54,11 @@ struct FASTDEPLOY_DECL ClassifyResult : public BaseResult {
   /// Clear result
   void Clear();
 
+  /// Copy constructor
+  ClassifyResult(const ClassifyResult& other) = default;
+  /// Move assignment
+  ClassifyResult& operator=(ClassifyResult&& other);
+
   /// Debug function, convert the result to string to print
   std::string Str();
 };

python/fastdeploy/vision/classification/__init__.py

@@ -14,8 +14,9 @@
 from __future__ import absolute_import
 
 from .contrib.yolov5cls import YOLOv5Cls
-from .ppcls import PaddleClasModel
+from .ppcls import *
 from .contrib.resnet import ResNet
+
 PPLCNet = PaddleClasModel
 PPLCNetv2 = PaddleClasModel
 EfficientNet = PaddleClasModel

python/fastdeploy/vision/classification/ppcls/__init__.py

@@ -18,6 +18,42 @@ from .... import FastDeployModel, ModelFormat
 from .... import c_lib_wrap as C
 
 
+class PaddleClasPreprocessor:
+    def __init__(self, config_file):
+        """Create a preprocessor for PaddleClasModel from configuration file
+
+        :param config_file: (str)Path of configuration file, e.g resnet50/inference_cls.yaml
+        """
+        self._preprocessor = C.vision.classification.PaddleClasPreprocessor(
+            config_file)
+
+    def run(self, input_ims):
+        """Preprocess input images for PaddleClasModel
+
+        :param: input_ims: (list of numpy.ndarray)The input image
+        :return: list of FDTensor
+        """
+        return self._preprocessor.run(input_ims)
+
+
+class PaddleClasPostprocessor:
+    def __init__(self, topk=1):
+        """Create a postprocessor for PaddleClasModel
+
+        :param topk: (int)Filter the top k classify label
+        """
+        self._postprocessor = C.vision.classification.PaddleClasPostprocessor(
+            topk)
+
+    def run(self, runtime_results):
+        """Postprocess the runtime results for PaddleClasModel
+
+        :param: runtime_results: (list of FDTensor)The output FDTensor results from runtime
+        :return: list of ClassifyResult(If the runtime_results is predict by batched samples, the length of this list equals to the batch size)
+        """
+        return self._postprocessor.run(runtime_results)
+
+
 class PaddleClasModel(FastDeployModel):
     def __init__(self,
                  model_file,
@@ -45,9 +81,35 @@ class PaddleClasModel(FastDeployModel):
     def predict(self, im, topk=1):
         """Classify an input image
 
-        :param im: (numpy.ndarray)The input image data, 3-D array with layout HWC, BGR format
+        :param im: (numpy.ndarray) The input image data, a 3-D array with layout HWC, BGR format
-        :param topk: (int)The topk result by the classify confidence score, default 1
+        :param topk: (int) Filter the topk classify result, default 1
         :return: ClassifyResult
         """
 
-        return self._model.predict(im, topk)
+        self.postprocessor.topk = topk
+        return self._model.predict(im)
+
+    def batch_predict(self, images):
+        """Classify a batch of input image
+
+        :param im: (list of numpy.ndarray) The input image list, each element is a 3-D array with layout HWC, BGR format
+        :return list of ClassifyResult
+        """
+
+        return self._model.batch_predict(images)
+
+    @property
+    def preprocessor(self):
+        """Get PaddleClasPreprocessor object of the loaded model
+
+        :return PaddleClasPreprocessor
+        """
+        return self._model.preprocessor
+
+    @property
+    def postprocessor(self):
+        """Get PaddleClasPostprocessor object of the loaded model
+
+        :return PaddleClasPostprocessor
+        """
+        return self._model.postprocessor

tests/models/test_mobilenetv2.py

@@ -22,9 +22,11 @@ import runtime_config as rc
 
 def test_classification_mobilenetv2():
     model_url = "https://bj.bcebos.com/paddlehub/fastdeploy/MobileNetV1_x0_25_infer.tgz"
-    input_url = "https://gitee.com/paddlepaddle/PaddleClas/raw/release/2.4/deploy/images/ImageNet/ILSVRC2012_val_00000010.jpeg"
+    input_url1 = "https://gitee.com/paddlepaddle/PaddleClas/raw/release/2.4/deploy/images/ImageNet/ILSVRC2012_val_00000010.jpeg"
+    input_url2 = "https://gitee.com/paddlepaddle/PaddleClas/raw/release/2.4/deploy/images/ImageNet/ILSVRC2012_val_00030010.jpeg"
     fd.download_and_decompress(model_url, "resources")
-    fd.download(input_url, "resources")
+    fd.download(input_url1, "resources")
+    fd.download(input_url2, "resources")
     model_path = "resources/MobileNetV1_x0_25_infer"
 
     model_file = "resources/MobileNetV1_x0_25_infer/inference.pdmodel"
@@ -33,18 +35,67 @@ def test_classification_mobilenetv2():
     model = fd.vision.classification.PaddleClasModel(
         model_file, params_file, config_file, runtime_option=rc.test_option)
 
-    expected_label_ids = [153, 333, 259, 338, 265, 154]
+    expected_label_ids_1 = [153, 333, 259, 338, 265, 154]
-    expected_scores = [
+    expected_scores_1 = [
         0.221088, 0.109457, 0.078668, 0.076814, 0.052401, 0.048206
     ]
+    expected_label_ids_2 = [80, 23, 93, 99, 143, 7]
+    expected_scores_2 = [
+        0.975599, 0.014083, 0.003821, 0.001571, 0.001233, 0.000924
+    ]
+
     # compare diff
-    im = cv2.imread("./resources/ILSVRC2012_val_00000010.jpeg")
+    im1 = cv2.imread("./resources/ILSVRC2012_val_00000010.jpeg")
-    for i in range(2):
+    im2 = cv2.imread("./resources/ILSVRC2012_val_00030010.jpeg")
-        result = model.predict(im, topk=6)
+
-        diff_label = np.fabs(
+    #    for i in range(3000000):
-            np.array(result.label_ids) - np.array(expected_label_ids))
+    while True:
-        diff_scores = np.fabs(
+        # test single predict
-            np.array(result.scores) - np.array(expected_scores))
+        model.postprocessor.topk = 6
-        assert diff_label.max() < 1e-06, "There's difference in classify label."
+        result1 = model.predict(im1)
-        assert diff_scores.max(
+        result2 = model.predict(im2)
-        ) < 1e-05, "There's difference in classify score."
+
+        diff_label_1 = np.fabs(
+            np.array(result1.label_ids) - np.array(expected_label_ids_1))
+        diff_label_2 = np.fabs(
+            np.array(result2.label_ids) - np.array(expected_label_ids_2))
+
+        diff_scores_1 = np.fabs(
+            np.array(result1.scores) - np.array(expected_scores_1))
+        diff_scores_2 = np.fabs(
+            np.array(result2.scores) - np.array(expected_scores_2))
+        assert diff_label_1.max(
+        ) < 1e-06, "There's difference in classify label 1."
+        assert diff_scores_1.max(
+        ) < 1e-05, "There's difference in classify score 1."
+        assert diff_label_2.max(
+        ) < 1e-06, "There's difference in classify label 2."
+        assert diff_scores_2.max(
+        ) < 1e-05, "There's difference in classify score 2."
+
+        # test batch predict
+        results = model.batch_predict([im1, im2])
+        result1 = results[0]
+        result2 = results[1]
+
+        diff_label_1 = np.fabs(
+            np.array(result1.label_ids) - np.array(expected_label_ids_1))
+        diff_label_2 = np.fabs(
+            np.array(result2.label_ids) - np.array(expected_label_ids_2))
+
+        diff_scores_1 = np.fabs(
+            np.array(result1.scores) - np.array(expected_scores_1))
+        diff_scores_2 = np.fabs(
+            np.array(result2.scores) - np.array(expected_scores_2))
+        assert diff_label_1.max(
+        ) < 1e-06, "There's difference in classify label 1."
+        assert diff_scores_1.max(
+        ) < 1e-05, "There's difference in classify score 1."
+        assert diff_label_2.max(
+        ) < 1e-06, "There's difference in classify label 2."
+        assert diff_scores_2.max(
+        ) < 1e-05, "There's difference in classify score 2."
+
+
+if __name__ == "__main__":
+    test_classification_mobilenetv2()