Merge branch 'develop' into ce

core/general-client/include/general_model.h

@@ -49,6 +49,8 @@ class ModelRes {
                             res._int64_value_map.end());
     _float_value_map.insert(res._float_value_map.begin(),
                             res._float_value_map.end());
+    _int32_value_map.insert(res._int32_value_map.begin(),
+                            res._int32_value_map.end());
     _shape_map.insert(res._shape_map.begin(), res._shape_map.end());
     _lod_map.insert(res._lod_map.begin(), res._lod_map.end());
   }
@@ -60,6 +62,9 @@ class ModelRes {
     _float_value_map.insert(
         std::make_move_iterator(std::begin(res._float_value_map)),
         std::make_move_iterator(std::end(res._float_value_map)));
+    _int32_value_map.insert(
+        std::make_move_iterator(std::begin(res._int32_value_map)),
+        std::make_move_iterator(std::end(res._int32_value_map)));
     _shape_map.insert(std::make_move_iterator(std::begin(res._shape_map)),
                       std::make_move_iterator(std::end(res._shape_map)));
     _lod_map.insert(std::make_move_iterator(std::begin(res._lod_map)),
@@ -78,6 +83,12 @@ class ModelRes {
   std::vector<float>&& get_float_by_name_with_rv(const std::string& name) {
     return std::move(_float_value_map[name]);
   }
+  const std::vector<int32_t>& get_int32_by_name(const std::string& name) {
+    return _int32_value_map[name];
+  }
+  std::vector<int32_t>&& get_int32_by_name_with_rv(const std::string& name) {
+    return std::move(_int32_value_map[name]);
+  }
   const std::vector<int>& get_shape_by_name(const std::string& name) {
     return _shape_map[name];
   }
@@ -103,6 +114,9 @@ class ModelRes {
       _float_value_map.insert(
           std::make_move_iterator(std::begin(res._float_value_map)),
           std::make_move_iterator(std::end(res._float_value_map)));
+      _int32_value_map.insert(
+          std::make_move_iterator(std::begin(res._int32_value_map)),
+          std::make_move_iterator(std::end(res._int32_value_map)));
       _shape_map.insert(std::make_move_iterator(std::begin(res._shape_map)),
                         std::make_move_iterator(std::end(res._shape_map)));
       _lod_map.insert(std::make_move_iterator(std::begin(res._lod_map)),
@@ -115,6 +129,7 @@ class ModelRes {
   std::string _engine_name;
   std::map<std::string, std::vector<int64_t>> _int64_value_map;
   std::map<std::string, std::vector<float>> _float_value_map;
+  std::map<std::string, std::vector<int32_t>> _int32_value_map;
   std::map<std::string, std::vector<int>> _shape_map;
   std::map<std::string, std::vector<int>> _lod_map;
 };
@@ -145,6 +160,14 @@ class PredictorRes {
                                                  const std::string& name) {
     return std::move(_models[model_idx].get_float_by_name_with_rv(name));
   }
+  const std::vector<int32_t>& get_int32_by_name(const int model_idx,
+                                                const std::string& name) {
+    return _models[model_idx].get_int32_by_name(name);
+  }
+  std::vector<int32_t>&& get_int32_by_name_with_rv(const int model_idx,
+                                                   const std::string& name) {
+    return std::move(_models[model_idx].get_int32_by_name_with_rv(name));
+  }
   const std::vector<int>& get_shape_by_name(const int model_idx,
                                             const std::string& name) {
     return _models[model_idx].get_shape_by_name(name);

core/general-client/src/general_model.cpp

@@ -207,17 +207,28 @@ int PredictorClient::batch_predict(
     for (auto &name : int_feed_name) {
       int idx = _feed_name_to_idx[name];
       Tensor *tensor = tensor_vec[idx];
-      VLOG(2) << "prepare int feed " << name << " shape size "
-              << int_shape[vec_idx].size();
+      if (_type[idx] == 0) {
+        VLOG(2) << "prepare int64 feed " << name << " shape size "
+                << int_shape[vec_idx].size();
+        VLOG(3) << "feed var name " << name << " index " << vec_idx
+                << "first data " << int_feed[vec_idx][0];
+        for (uint32_t j = 0; j < int_feed[vec_idx].size(); ++j) {
+          tensor->add_int64_data(int_feed[vec_idx][j]);
+        }
+      } else if (_type[idx] == 2) {
+        VLOG(2) << "prepare int32 feed " << name << " shape size "
+                << int_shape[vec_idx].size();
+        VLOG(3) << "feed var name " << name << " index " << vec_idx
+                << "first data " << int32_t(int_feed[vec_idx][0]);
+        for (uint32_t j = 0; j < int_feed[vec_idx].size(); ++j) {
+          tensor->add_int_data(int32_t(int_feed[vec_idx][j]));
+        }
+      }
+
       for (uint32_t j = 0; j < int_shape[vec_idx].size(); ++j) {
         tensor->add_shape(int_shape[vec_idx][j]);
       }
-      tensor->set_elem_type(0);
-      VLOG(3) << "feed var name " << name << " index " << vec_idx
-              << "first data " << int_feed[vec_idx][0];
-      for (uint32_t j = 0; j < int_feed[vec_idx].size(); ++j) {
-        tensor->add_int64_data(int_feed[vec_idx][j]);
-      }
+      tensor->set_elem_type(_type[idx]);
       vec_idx++;
     }
 
@@ -284,18 +295,25 @@ int PredictorClient::batch_predict(
       for (auto &name : fetch_name) {
         // int idx = _fetch_name_to_idx[name];
         if (_fetch_name_to_type[name] == 0) {
-          VLOG(2) << "ferch var " << name << "type int";
+          VLOG(2) << "ferch var " << name << "type int64";
           int size = output.insts(0).tensor_array(idx).int64_data_size();
           model._int64_value_map[name] = std::vector<int64_t>(
               output.insts(0).tensor_array(idx).int64_data().begin(),
               output.insts(0).tensor_array(idx).int64_data().begin() + size);
-        } else {
+        } else if (_fetch_name_to_type[name] == 1) {
           VLOG(2) << "fetch var " << name << "type float";
           int size = output.insts(0).tensor_array(idx).float_data_size();
           model._float_value_map[name] = std::vector<float>(
               output.insts(0).tensor_array(idx).float_data().begin(),
               output.insts(0).tensor_array(idx).float_data().begin() + size);
+        } else if (_fetch_name_to_type[name] == 2) {
+          VLOG(2) << "fetch var " << name << "type int32";
+          int size = output.insts(0).tensor_array(idx).int_data_size();
+          model._int32_value_map[name] = std::vector<int32_t>(
+              output.insts(0).tensor_array(idx).int_data().begin(),
+              output.insts(0).tensor_array(idx).int_data().begin() + size);
         }
+
         idx += 1;
       }
       predict_res_batch.add_model_res(std::move(model));
@@ -442,12 +460,19 @@ int PredictorClient::numpy_predict(
     for (auto &name : int_feed_name) {
       int idx = _feed_name_to_idx[name];
       Tensor *tensor = tensor_vec[idx];
-      VLOG(2) << "prepare int feed " << name << " shape size "
-              << int_shape[vec_idx].size();
+
       for (uint32_t j = 0; j < int_shape[vec_idx].size(); ++j) {
         tensor->add_shape(int_shape[vec_idx][j]);
       }
-      tensor->set_elem_type(0);
+      tensor->set_elem_type(_type[idx]);
+
+      if (_type[idx] == 0) {
+        VLOG(2) << "prepare int feed " << name << " shape size "
+                << int_shape[vec_idx].size();
+      } else {
+        VLOG(2) << "prepare int32 feed " << name << " shape size "
+                << int_shape[vec_idx].size();
+      }
 
       const int int_shape_size = int_shape[vec_idx].size();
       switch (int_shape_size) {
@@ -457,7 +482,11 @@ int PredictorClient::numpy_predict(
             for (ssize_t j = 0; j < int_array.shape(1); j++) {
               for (ssize_t k = 0; k < int_array.shape(2); k++) {
                 for (ssize_t l = 0; k < int_array.shape(3); l++) {
-                  tensor->add_int64_data(int_array(i, j, k, l));
+                  if (_type[idx] == 0) {
+                    tensor->add_int64_data(int_array(i, j, k, l));
+                  } else {
+                    tensor->add_int_data(int_array(i, j, k, l));
+                  }
                 }
               }
             }
@@ -469,7 +498,11 @@ int PredictorClient::numpy_predict(
           for (ssize_t i = 0; i < int_array.shape(0); i++) {
             for (ssize_t j = 0; j < int_array.shape(1); j++) {
               for (ssize_t k = 0; k < int_array.shape(2); k++) {
-                tensor->add_int64_data(int_array(i, j, k));
+                if (_type[idx] == 0) {
+                  tensor->add_int64_data(int_array(i, j, k));
+                } else {
+                  tensor->add_int_data(int_array(i, j, k));
+                }
               }
             }
           }
@@ -479,7 +512,11 @@ int PredictorClient::numpy_predict(
           auto int_array = int_feed[vec_idx].unchecked<2>();
           for (ssize_t i = 0; i < int_array.shape(0); i++) {
             for (ssize_t j = 0; j < int_array.shape(1); j++) {
-              tensor->add_int64_data(int_array(i, j));
+              if (_type[idx] == 0) {
+                tensor->add_int64_data(int_array(i, j));
+              } else {
+                tensor->add_int_data(int_array(i, j));
+              }
             }
           }
           break;
@@ -487,7 +524,11 @@ int PredictorClient::numpy_predict(
         case 1: {
           auto int_array = int_feed[vec_idx].unchecked<1>();
           for (ssize_t i = 0; i < int_array.shape(0); i++) {
-            tensor->add_int64_data(int_array(i));
+            if (_type[idx] == 0) {
+              tensor->add_int64_data(int_array(i));
+            } else {
+              tensor->add_int_data(int_array(i));
+            }
           }
           break;
         }
@@ -557,17 +598,23 @@ int PredictorClient::numpy_predict(
       for (auto &name : fetch_name) {
         // int idx = _fetch_name_to_idx[name];
         if (_fetch_name_to_type[name] == 0) {
-          VLOG(2) << "ferch var " << name << "type int";
+          VLOG(2) << "ferch var " << name << "type int64";
           int size = output.insts(0).tensor_array(idx).int64_data_size();
           model._int64_value_map[name] = std::vector<int64_t>(
               output.insts(0).tensor_array(idx).int64_data().begin(),
               output.insts(0).tensor_array(idx).int64_data().begin() + size);
-        } else {
+        } else if (_fetch_name_to_type[name] == 1) {
           VLOG(2) << "fetch var " << name << "type float";
           int size = output.insts(0).tensor_array(idx).float_data_size();
           model._float_value_map[name] = std::vector<float>(
               output.insts(0).tensor_array(idx).float_data().begin(),
               output.insts(0).tensor_array(idx).float_data().begin() + size);
+        } else if (_fetch_name_to_type[name] == 2) {
+          VLOG(2) << "fetch var " << name << "type int32";
+          int size = output.insts(0).tensor_array(idx).int_data_size();
+          model._int32_value_map[name] = std::vector<int32_t>(
+              output.insts(0).tensor_array(idx).int_data().begin(),
+              output.insts(0).tensor_array(idx).int_data().begin() + size);
         }
         idx += 1;
       }
@@ -601,7 +648,6 @@ int PredictorClient::numpy_predict(
   _api.thrd_clear();
   return 0;
 }
-
 }  // namespace general_model
 }  // namespace paddle_serving
 }  // namespace baidu

core/general-server/op/general_reader_op.cpp

@@ -126,9 +126,12 @@ int GeneralReaderOp::inference() {
     if (elem_type[i] == 0) {  // int64
       elem_size[i] = sizeof(int64_t);
       lod_tensor.dtype = paddle::PaddleDType::INT64;
-    } else {
+    } else if (elem_type[i] == 1) {
       elem_size[i] = sizeof(float);
       lod_tensor.dtype = paddle::PaddleDType::FLOAT32;
+    } else if (elem_type[i] == 2) {
+      elem_size[i] = sizeof(int32_t);
+      lod_tensor.dtype = paddle::PaddleDType::INT32;
     }
 
     if (model_config->_is_lod_feed[i]) {
@@ -159,8 +162,10 @@ int GeneralReaderOp::inference() {
         int data_len = 0;
         if (tensor.int64_data_size() > 0) {
           data_len = tensor.int64_data_size();
-        } else {
+        } else if (tensor.float_data_size() > 0) {
           data_len = tensor.float_data_size();
+        } else if (tensor.int_data_size() > 0) {
+          data_len = tensor.int_data_size();
         }
         VLOG(2) << "tensor size for var[" << i << "]: " << data_len;
         tensor_size += data_len;
@@ -198,6 +203,8 @@ int GeneralReaderOp::inference() {
   for (int i = 0; i < var_num; ++i) {
     if (elem_type[i] == 0) {
       int64_t *dst_ptr = static_cast<int64_t *>(out->at(i).data.data());
+      VLOG(2) << "first element data in var[" << i << "] is "
+              << req->insts(0).tensor_array(i).int64_data(0);
       int offset = 0;
       for (int j = 0; j < batch_size; ++j) {
         int elem_num = req->insts(j).tensor_array(i).int64_data_size();
@@ -210,8 +217,10 @@ int GeneralReaderOp::inference() {
           offset += capacity[i];
         }
       }
-    } else {
+    } else if (elem_type[i] == 1) {
       float *dst_ptr = static_cast<float *>(out->at(i).data.data());
+      VLOG(2) << "first element data in var[" << i << "] is "
+              << req->insts(0).tensor_array(i).float_data(0);
       int offset = 0;
       for (int j = 0; j < batch_size; ++j) {
         int elem_num = req->insts(j).tensor_array(i).float_data_size();
@@ -224,6 +233,22 @@ int GeneralReaderOp::inference() {
           offset += capacity[i];
         }
       }
+    } else if (elem_type[i] == 2) {
+      int32_t *dst_ptr = static_cast<int32_t *>(out->at(i).data.data());
+      VLOG(2) << "first element data in var[" << i << "] is "
+              << req->insts(0).tensor_array(i).int_data(0);
+      int offset = 0;
+      for (int j = 0; j < batch_size; ++j) {
+        int elem_num = req->insts(j).tensor_array(i).int_data_size();
+        for (int k = 0; k < elem_num; ++k) {
+          dst_ptr[offset + k] = req->insts(j).tensor_array(i).int_data(k);
+        }
+        if (out->at(i).lod.size() == 1) {
+          offset = out->at(i).lod[0][j + 1];
+        } else {
+          offset += capacity[i];
+        }
+      }
     }
   }
 

core/general-server/op/general_response_op.cpp

@@ -91,7 +91,6 @@ int GeneralResponseOp::inference() {
 
     for (auto &idx : fetch_index) {
       Tensor *tensor = fetch_inst->add_tensor_array();
-      tensor->set_elem_type(1);
       if (model_config->_is_lod_fetch[idx]) {
         VLOG(2) << "out[" << idx << "] " << model_config->_fetch_name[idx]
                 << " is lod_tensor";
@@ -116,7 +115,7 @@ int GeneralResponseOp::inference() {
         cap *= in->at(idx).shape[j];
       }
       if (in->at(idx).dtype == paddle::PaddleDType::INT64) {
-        VLOG(2) << "Prepare float var [" << model_config->_fetch_name[idx]
+        VLOG(2) << "Prepare int64 var [" << model_config->_fetch_name[idx]
                 << "].";
         int64_t *data_ptr = static_cast<int64_t *>(in->at(idx).data.data());
         if (model_config->_is_lod_fetch[idx]) {
@@ -157,6 +156,27 @@ int GeneralResponseOp::inference() {
         }
         VLOG(2) << "fetch var [" << model_config->_fetch_name[idx] << "] ready";
         var_idx++;
+      } else if (in->at(idx).dtype == paddle::PaddleDType::INT32) {
+        VLOG(2) << "Prepare int32 var [" << model_config->_fetch_name[idx]
+                << "].";
+        int32_t *data_ptr = static_cast<int32_t *>(in->at(idx).data.data());
+        if (model_config->_is_lod_fetch[idx]) {
+          FetchInst *fetch_p = output->mutable_insts(0);
+          for (int j = 0; j < in->at(idx).lod[0].size(); ++j) {
+            fetch_p->mutable_tensor_array(var_idx)->add_lod(
+                in->at(idx).lod[0][j]);
+          }
+          for (int j = 0; j < cap; ++j) {
+            fetch_p->mutable_tensor_array(var_idx)->add_int_data(data_ptr[j]);
+          }
+        } else {
+          FetchInst *fetch_p = output->mutable_insts(0);
+          for (int j = 0; j < cap; ++j) {
+            fetch_p->mutable_tensor_array(var_idx)->add_int_data(data_ptr[j]);
+          }
+        }
+        VLOG(2) << "fetch var [" << model_config->_fetch_name[idx] << "] ready";
+        var_idx++;
       }
     }
   }

core/predictor/framework/infer.h

@@ -603,13 +603,13 @@ class VersionedInferEngine : public InferEngine {
       LOG(ERROR) << "Failed generate engine with type:" << engine_type;
       return -1;
     }
-    VLOG(2) << "FLGS_logtostderr " << FLAGS_logtostderr;
+    VLOG(2) << "FLAGS_logtostderr " << FLAGS_logtostderr;
     int tmp = FLAGS_logtostderr;
     if (engine->proc_initialize(conf, version) != 0) {
       LOG(ERROR) << "Failed initialize engine, type:" << engine_type;
       return -1;
     }
-    VLOG(2) << "FLGS_logtostderr " << FLAGS_logtostderr;
+    VLOG(2) << "FLAGS_logtostderr " << FLAGS_logtostderr;
     FLAGS_logtostderr = tmp;
     auto r = _versions.insert(std::make_pair(engine->version(), engine));
     if (!r.second) {

python/examples/yolov4/README.md

+# Yolov4 Detection Service
+
+([简体中文](README_CN.md)|English)
+
+## Get Model
+
+```
+python -m paddle_serving_app.package --get_model yolov4
+tar -xzvf yolov4.tar.gz
+```
+
+## Start RPC Service
+
+```
+python -m paddle_serving_server_gpu.serve --model yolov4_model --port 9393 --gpu_ids 0
+```
+
+## Prediction
+
+```
+python test_client.py 000000570688.jpg
+```
+After the prediction is completed, a json file to save the prediction result and a picture with the detection result box will be generated in the `./outpu folder.

python/examples/yolov4/README_CN.md

+# Yolov4 检测服务
+
+(简体中文|[English](README.md))
+
+## 获取模型
+
+```
+python -m paddle_serving_app.package --get_model yolov4
+tar -xzvf yolov4.tar.gz
+```
+
+## 启动RPC服务
+
+```
+python -m paddle_serving_server_gpu.serve --model yolov4_model --port 9393 --gpu_ids 0
+```
+
+## 预测
+
+```
+python test_client.py 000000570688.jpg
+```
+
+预测完成会在`./output`文件夹下生成保存预测结果的json文件以及标出检测结果框的图片。

python/examples/yolov4/label_list.txt

+person
+bicycle
+car
+motorcycle
+airplane
+bus
+train
+truck
+boat
+traffic light
+fire hydrant
+stop sign
+parking meter
+bench
+bird
+cat
+dog
+horse
+sheep
+cow
+elephant
+bear
+zebra
+giraffe
+backpack
+umbrella
+handbag
+tie
+suitcase
+frisbee
+skis
+snowboard
+sports ball
+kite
+baseball bat
+baseball glove
+skateboard
+surfboard
+tennis racket
+bottle
+wine glass
+cup
+fork
+knife
+spoon
+bowl
+banana
+apple
+sandwich
+orange
+broccoli
+carrot
+hot dog
+pizza
+donut
+cake
+chair
+couch
+potted plant
+bed
+dining table
+toilet
+tv
+laptop
+mouse
+remote
+keyboard
+cell phone
+microwave
+oven
+toaster
+sink
+refrigerator
+book
+clock
+vase
+scissors
+teddy bear
+hair drier
+toothbrush

python/examples/yolov4/test_client.py

+# Copyright (c) 2020 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 sys
+import numpy as np
+from paddle_serving_client import Client
+from paddle_serving_app.reader import *
+import cv2
+preprocess = Sequential([
+    File2Image(), BGR2RGB(), Resize(
+        (608, 608), interpolation=cv2.INTER_LINEAR), Div(255.0), Transpose(
+            (2, 0, 1))
+])
+
+postprocess = RCNNPostprocess("label_list.txt", "output", [608, 608])
+client = Client()
+
+client.load_client_config("yolov4_client/serving_client_conf.prototxt")
+client.connect(['127.0.0.1:9393'])
+
+im = preprocess(sys.argv[1])
+print(im.shape)
+fetch_map = client.predict(
+    feed={
+        "image": im,
+        "im_size": np.array(list(im.shape[1:])),
+    },
+    fetch=["save_infer_model/scale_0.tmp_0"])
+fetch_map["image"] = sys.argv[1]
+postprocess(fetch_map)

python/paddle_serving_app/models/model_list.py

@@ -24,7 +24,7 @@ class ServingModels(object):
             "SentimentAnalysis"] = ["senta_bilstm", "senta_bow", "senta_cnn"]
         self.model_dict["SemanticRepresentation"] = ["ernie"]
         self.model_dict["ChineseWordSegmentation"] = ["lac"]
-        self.model_dict["ObjectDetection"] = ["faster_rcnn"]
+        self.model_dict["ObjectDetection"] = ["faster_rcnn", "yolov4"]
         self.model_dict["ImageSegmentation"] = [
             "unet", "deeplabv3", "deeplabv3+cityscapes"
         ]

python/paddle_serving_app/reader/image_reader.py

@@ -280,10 +280,11 @@ class SegPostprocess(object):
 
 
 class RCNNPostprocess(object):
-    def __init__(self, label_file, output_dir):
+    def __init__(self, label_file, output_dir, resize_shape=None):
         self.output_dir = output_dir
         self.label_file = label_file
         self.label_list = []
+        self.resize_shape = resize_shape
         with open(label_file) as fin:
             for line in fin:
                 self.label_list.append(line.strip())
@@ -378,6 +379,13 @@ class RCNNPostprocess(object):
             xmax = xmin + w
             ymax = ymin + h
 
+            img_w, img_h = image.size
+            if self.resize_shape is not None:
+                xmin = xmin * img_w / self.resize_shape[0]
+                xmax = xmax * img_w / self.resize_shape[0]
+                ymin = ymin * img_h / self.resize_shape[1]
+                ymax = ymax * img_h / self.resize_shape[1]
+
             color = tuple(color_list[catid])
 
             # draw bbox

python/paddle_serving_client/__init__.py

@@ -28,8 +28,11 @@ sys.path.append(
     os.path.join(os.path.abspath(os.path.dirname(__file__)), 'proto'))
 from .proto import multi_lang_general_model_service_pb2_grpc
 
-int_type = 0
-float_type = 1
+int64_type = 0
+float32_type = 1
+int32_type = 2
+int_type = set([int64_type, int32_type])
+float_type = set([float32_type])
 
 
 class _NOPProfiler(object):
@@ -279,7 +282,7 @@ class Client(object):
                     raise ValueError("Wrong feed name: {}.".format(key))
                 #if not isinstance(feed_i[key], np.ndarray):
                 self.shape_check(feed_i, key)
-                if self.feed_types_[key] == int_type:
+                if self.feed_types_[key] in int_type:
                     if i == 0:
                         int_feed_names.append(key)
                         if isinstance(feed_i[key], np.ndarray):
@@ -292,7 +295,7 @@ class Client(object):
                     else:
                         int_slot.append(feed_i[key])
                         self.all_numpy_input = False
-                elif self.feed_types_[key] == float_type:
+                elif self.feed_types_[key] in float_type:
                     if i == 0:
                         float_feed_names.append(key)
                         if isinstance(feed_i[key], np.ndarray):
@@ -339,7 +342,7 @@ class Client(object):
             result_map = {}
             # result map needs to be a numpy array
             for i, name in enumerate(fetch_names):
-                if self.fetch_names_to_type_[name] == int_type:
+                if self.fetch_names_to_type_[name] == int64_type:
                     # result_map[name] will be py::array(numpy array)
                     result_map[name] = result_batch_handle.get_int64_by_name(
                         mi, name)
@@ -348,7 +351,7 @@ class Client(object):
                     if name in self.lod_tensor_set:
                         result_map["{}.lod".format(
                             name)] = result_batch_handle.get_lod(mi, name)
-                elif self.fetch_names_to_type_[name] == float_type:
+                elif self.fetch_names_to_type_[name] == float32_type:
                     result_map[name] = result_batch_handle.get_float_by_name(
                         mi, name)
                     shape = result_batch_handle.get_shape(mi, name)
@@ -356,6 +359,16 @@ class Client(object):
                     if name in self.lod_tensor_set:
                         result_map["{}.lod".format(
                             name)] = result_batch_handle.get_lod(mi, name)
+
+                elif self.fetch_names_to_type_[name] == int32_type:
+                    # result_map[name] will be py::array(numpy array)
+                    result_map[name] = result_batch_handle.get_int32_by_name(
+                        mi, name)
+                    shape = result_batch_handle.get_shape(mi, name)
+                    result_map[name].shape = shape
+                    if name in self.lod_tensor_set:
+                        result_map["{}.lod".format(
+                            name)] = result_batch_handle.get_lod(mi, name)
             multi_result_map.append(result_map)
         ret = None
         if len(model_engine_names) == 1:
@@ -460,6 +473,8 @@ class MultiLangClient(object):
                             data = np.array(var, dtype="int64")
                         elif v_type == 1:  # float32
                             data = np.array(var, dtype="float32")
+                        elif v_type == 2:  #int32
+                            data = np.array(var, dtype="int32")
                         else:
                             raise Exception("error type.")
                     else:
@@ -478,6 +493,11 @@ class MultiLangClient(object):
                             tensor.float_data.extend(var.reshape(-1).tolist())
                         else:
                             tensor.float_data.extend(self._flatten_list(var))
+                    elif v_type == 2:  #int32
+                        if isinstance(car, np.array):
+                            tensor.int_data.extend(var.reshape(-1).tolist())
+                        else:
+                            tensor.int_data.extend(self._flatten_list(var))
                     else:
                         raise Exception("error type.")
                 if isinstance(var, np.ndarray):
@@ -509,6 +529,9 @@ class MultiLangClient(object):
                 elif v_type == 1:  # float32
                     result_map[name] = np.array(
                         list(var.float_data), dtype="float32")
+                elif v_type == 2:  # int32
+                    result_map[name] = np.array(
+                        list(var.int_data), dtype="int32")
                 else:
                     raise Exception("error type.")
             result_map[name].shape = list(var.shape)

python/paddle_serving_client/io/__init__.py

@@ -48,16 +48,18 @@ def save_model(server_model_folder,
 
     config = model_conf.GeneralModelConfig()
 
+    #int64 = 0; float32 = 1; int32 = 2;
     for key in feed_var_dict:
         feed_var = model_conf.FeedVar()
         feed_var.alias_name = key
         feed_var.name = feed_var_dict[key].name
         feed_var.is_lod_tensor = feed_var_dict[key].lod_level >= 1
-        if feed_var_dict[key].dtype == core.VarDesc.VarType.INT32 or \
-           feed_var_dict[key].dtype == core.VarDesc.VarType.INT64:
+        if feed_var_dict[key].dtype == core.VarDesc.VarType.INT64:
             feed_var.feed_type = 0
         if feed_var_dict[key].dtype == core.VarDesc.VarType.FP32:
             feed_var.feed_type = 1
+        if feed_var_dict[key].dtype == core.VarDesc.VarType.INT32:
+            feed_var.feed_type = 2
         if feed_var.is_lod_tensor:
             feed_var.shape.extend([-1])
         else:
@@ -73,13 +75,12 @@ def save_model(server_model_folder,
         fetch_var.alias_name = key
         fetch_var.name = fetch_var_dict[key].name
         fetch_var.is_lod_tensor = fetch_var_dict[key].lod_level >= 1
-        if fetch_var_dict[key].dtype == core.VarDesc.VarType.INT32 or \
-           fetch_var_dict[key].dtype == core.VarDesc.VarType.INT64:
+        if fetch_var_dict[key].dtype == core.VarDesc.VarType.INT64:
             fetch_var.fetch_type = 0
-
         if fetch_var_dict[key].dtype == core.VarDesc.VarType.FP32:
             fetch_var.fetch_type = 1
-
+        if fetch_var_dict[key].dtype == core.VarDesc.VarType.INT32:
+            fetch_var.fetch_type = 2
         if fetch_var.is_lod_tensor:
             fetch_var.shape.extend([-1])
         else:

tools/serving_build.sh

@@ -499,6 +499,41 @@ function python_test_lac() {
     cd ..
 }
 
+function python_test_yolov4(){
+    #pwd:/ Serving/python/examples
+    local TYPE=$1
+    export SERVING_BIN=${SERVING_WORKDIR}/build-server-${TYPE}/core/general-server/serving
+    cd yolov4
+    case $TYPE in
+        CPU)
+            python -m paddle_serving_app.package --get_model yolov4
+            tar -xzvf yolov4.tar.gz
+            check_cmd "python -m paddle_serving_server.serve --model yolov4_model/ --port 9393 &"
+            sleep 5
+            check_cmd "python test_client.py 000000570688.jpg"
+            echo "yolov4 CPU RPC inference pass"
+            kill_server_process
+            ;;
+        GPU)
+            python -m paddle_serving_app.package --get_model yolov4
+            tar -xzvf yolov4.tar.gz
+            check_cmd "python -m paddle_serving_server_gpu.serve --model yolov4_model/ --port 9393 --gpu_ids 0 &"
+            sleep 5
+            check_cmd "python test_client.py 000000570688.jpg"
+            echo "yolov4 GPU RPC inference pass"
+            kill_server_process
+            ;;
+        *)
+            echo "error type"
+            exit 1
+            ;;
+    esac
+    echo "test yolov4 $TYPE finished as expected."
+    unset SERVING_BIN
+    cd ..
+}
+
+
 function python_run_test() {
     # Using the compiled binary
     local TYPE=$1 # pwd: /Serving
@@ -510,6 +545,7 @@ function python_run_test() {
     python_test_lac $TYPE # pwd: /Serving/python/examples
     python_test_multi_process $TYPE # pwd: /Serving/python/examples
     python_test_multi_fetch $TYPE # pwd: /Serving/python/examples
+    python_test_yolov4 $TYPE # pwd: /Serving/python/examples
     echo "test python $TYPE part finished as expected."
     cd ../.. # pwd: /Serving
 }