Merge pull request #123 from MRXLT/general-model-config

add  c++ code for general server

core/configure/proto/general_model_config.proto

@@ -15,12 +15,17 @@
 syntax = "proto2";
 package baidu.paddle_serving.configure;
 
-message Shape {
-  repeated int32 shape = 1;
-};
-
+message FeedVar {
+  required string name = 1;
+  required bool is_lod_tensor = 2;
+  required int32 feed_type = 3;
+  repeated int32 shape = 4;
+}
+message FetchVar {
+  required string name = 1;
+  repeated int32 shape = 2;
+}
 message GeneralModelConfig {
-  repeated bool is_lod_feed = 1;
-  repeated int32 feed_type = 2;
-  repeated Shape feed_shape = 4;
+  repeated FeedVar feed_var = 1;
+  repeated FetchVar fetch_var = 2;
 };

core/configure/proto/sdk_configure.proto

@@ -16,27 +16,27 @@ syntax = "proto2";
 package baidu.paddle_serving.configure;
 
 message ConnectionConf {
-  required int32 connect_timeout_ms = 1 [default = 2000];
-  required int32 rpc_timeout_ms = 2 [default = 20000];
-  required int32 connect_retry_count = 3 [default = 2];
-  required int32 max_connection_per_host = 4 [default = 100];
-  required int32 hedge_request_timeout_ms = 5 [default = -1];
-  required int32 hedge_fetch_retry_count = 6 [default = 2];
-  required string connection_type = 7 [default = "pooled"];
+  required int32 connect_timeout_ms = 1 [ default = 2000 ];
+  required int32 rpc_timeout_ms = 2 [ default = 20000 ];
+  required int32 connect_retry_count = 3 [ default = 2 ];
+  required int32 max_connection_per_host = 4 [ default = 100 ];
+  required int32 hedge_request_timeout_ms = 5 [ default = -1 ];
+  required int32 hedge_fetch_retry_count = 6 [ default = 2 ];
+  required string connection_type = 7 [ default = "pooled" ];
 };
 
 message NamingConf {
-  optional string cluster_filter_strategy = 1 [default="Default"];
-  optional string load_balance_strategy = 2 [default = "la"];
+  optional string cluster_filter_strategy = 1 [ default = "Default" ];
+  optional string load_balance_strategy = 2 [ default = "la" ];
   optional string cluster = 3;
 };
 
 message RpcParameter {
   // 0-NONE, 1-SNAPPY, 2-GZIP, 3-ZLIB, 4-LZ4
-  required int32 compress_type = 1 [default = 0];
-  required int32 package_size = 2 [default = 20];
-  required string protocol = 3 [default = "baidu_std"];
-  required int32 max_channel_per_request = 4 [default = 3];
+  required int32 compress_type = 1 [ default = 0 ];
+  required int32 package_size = 2 [ default = 20 ];
+  required string protocol = 3 [ default = "baidu_std" ];
+  required int32 max_channel_per_request = 4 [ default = 3 ];
 };
 
 message SplitConf {
@@ -53,12 +53,17 @@ message VariantConf {
   optional string variant_router = 6;
 };
 
-message WeightedRandomRenderConf { required string variant_weight_list = 1 [default= "50" ]; };
+message WeightedRandomRenderConf {
+  required string variant_weight_list = 1 [ default = "50" ];
+};
 
 message Predictor {
-  required string name = 1 [default="general_model"];
-  required string service_name = 2 [default="baidu.paddle_serving.predictor.general_model.GeneralModelService"];
-  required string endpoint_router = 3 [default="WeightedRandomRender"];
+  required string name = 1 [ default = "general_model" ];
+  required string service_name = 2 [
+    default =
+        "baidu.paddle_serving.predictor.general_model.GeneralModelService"
+  ];
+  required string endpoint_router = 3 [ default = "WeightedRandomRender" ];
   required WeightedRandomRenderConf weighted_random_render_conf = 4;
   repeated VariantConf variants = 5;
 };

core/configure/proto/server_configure.proto

@@ -53,6 +53,8 @@ message ResourceConf {
   required string model_toolkit_path = 1;
   required string model_toolkit_file = 2;
   optional string cube_config_file = 3;
+  optional string general_model_path = 4;
+  optional string general_model_file = 5;
 };
 
 // DAG node depency info

core/predictor/framework/resource.cpp

@@ -23,7 +23,6 @@ namespace predictor {
 
 using configure::ResourceConf;
 using configure::GeneralModelConfig;
-using configure::Shape;
 using rec::mcube::CubeAPI;
 // __thread bool p_thread_initialized = false;
 
@@ -47,15 +46,22 @@ std::shared_ptr<PaddleGeneralModelConfig> Resource::get_general_model_config() {
 }
 
 void Resource::print_general_model_config(
-    const std::shared_ptr<PaddleGeneralModelConfig> & config) {
+    const std::shared_ptr<PaddleGeneralModelConfig>& config) {
   if (config == nullptr) {
     LOG(INFO) << "paddle general model config is not set";
     return;
   }
-  LOG(INFO) << "Number of Feed Tensor: " << config->_feed_type.size();
+  LOG(INFO) << "Number of Feed Tensor: " << config->_feed_name.size();
   std::ostringstream oss;
+  LOG(INFO) << "Feed Name Info";
+  for (auto& feed_name : config->_feed_name) {
+    oss << feed_name << " ";
+  }
+  LOG(INFO) << oss.str();
+  oss.clear();
+  oss.str("");
   LOG(INFO) << "Feed Type Info";
-  for (auto & feed_type : config->_feed_type) {
+  for (auto& feed_type : config->_feed_type) {
     oss << feed_type << " ";
   }
   LOG(INFO) << oss.str();
@@ -71,7 +77,7 @@ void Resource::print_general_model_config(
   oss.clear();
   oss.str("");
   LOG(INFO) << "Capacity Info";
-  for (auto & cap : config->_capacity) {
+  for (auto& cap : config->_capacity) {
     oss << cap << " ";
   }
   LOG(INFO) << oss.str();
@@ -79,8 +85,8 @@ void Resource::print_general_model_config(
   oss.str("");
   LOG(INFO) << "Feed Shape Info";
   int tensor_idx = 0;
-  for (auto & shape : config->_feed_shape) {
-    for (auto & dim : shape) {
+  for (auto& shape : config->_feed_shape) {
+    for (auto& dim : shape) {
       oss << dim << " ";
     }
     LOG(INFO) << "Tensor[" << tensor_idx++ << "].shape: " << oss.str();
@@ -146,38 +152,69 @@ int Resource::initialize(const std::string& path, const std::string& file) {
   return 0;
 }
 
-int Resource::general_model_initialize(
-    const std::string& path, const std::string & file) {
+// model config
+int Resource::general_model_initialize(const std::string& path,
+                                       const std::string& file) {
   if (!FLAGS_enable_general_model) {
     return 0;
   }
+  ResourceConf resource_conf;
+  if (configure::read_proto_conf(path, file, &resource_conf) != 0) {
+    LOG(ERROR) << "Failed initialize resource from: " << path << "/" << file;
+    return -1;
+  }
+  int err = 0;
+  std::string general_model_path = resource_conf.general_model_path();
+  std::string general_model_file = resource_conf.general_model_file();
+  if (err != 0) {
+    LOG(ERROR) << "read general_model_path failed, path[" << path << "], file["
+               << file << "]";
+    return -1;
+  }
 
   GeneralModelConfig model_config;
-  if (configure::read_proto_conf(path, file, &model_config) != 0) {
-    LOG(ERROR) << "Failed initialize resource from: " << path << "/" << file;
+  if (configure::read_proto_conf(general_model_path.c_str(),
+                                 general_model_file.c_str(),
+                                 &model_config) != 0) {
+    LOG(ERROR) << "Failed initialize model config from: " << general_model_path
+               << "/" << general_model_file;
     return -1;
   }
 
   _config.reset(new PaddleGeneralModelConfig());
-  _config->_feed_type.resize(model_config.feed_type_size());
-  _config->_is_lod_feed.resize(model_config.is_lod_feed_size());
-  _config->_capacity.resize(model_config.feed_shape_size());
-  _config->_feed_shape.resize(model_config.feed_shape_size());
-  for (int i = 0; i < model_config.is_lod_feed_size(); ++i) {
-    _config->_feed_type[i] = model_config.feed_type(i);
-    if (model_config.is_lod_feed(i)) {
+  int feed_var_num = model_config.feed_var_size();
+  _config->_feed_name.resize(feed_var_num);
+  _config->_feed_type.resize(feed_var_num);
+  _config->_is_lod_feed.resize(feed_var_num);
+  _config->_capacity.resize(feed_var_num);
+  _config->_feed_shape.resize(feed_var_num);
+  for (int i = 0; i < feed_var_num; ++i) {
+    _config->_feed_name[i] = model_config.feed_var(i).name();
+    _config->_feed_type[i] = model_config.feed_var(i).feed_type();
+    if (model_config.feed_var(i).is_lod_tensor()) {
       _config->_feed_shape[i] = {-1};
       _config->_is_lod_feed[i] = true;
     } else {
       _config->_capacity[i] = 1;
       _config->_is_lod_feed[i] = false;
-      for (int j = 0; j < model_config.feed_shape(i).shape_size(); ++j) {
-        int dim = model_config.feed_shape(i).shape(j);
+      for (int j = 0; j < model_config.feed_var(i).shape_size(); ++j) {
+        int32_t dim = model_config.feed_var(i).shape(j);
         _config->_feed_shape[i].push_back(dim);
         _config->_capacity[i] *= dim;
       }
     }
   }
+
+  int fetch_var_num = model_config.fetch_var_size();
+  _config->_fetch_name.resize(fetch_var_num);
+  _config->_fetch_shape.resize(fetch_var_num);
+  for (int i = 0; i < fetch_var_num; ++i) {
+    _config->_fetch_name[i] = model_config.fetch_var(i).name();
+    for (int j = 0; j < model_config.fetch_var(i).shape_size(); ++j) {
+      int dim = model_config.fetch_var(i).shape(j);
+      _config->_fetch_shape[i].push_back(dim);
+    }
+  }
   return 0;
 }
 

core/predictor/framework/resource.h

@@ -33,16 +33,19 @@ class PaddleGeneralModelConfig {
   ~PaddleGeneralModelConfig() {}
 
  public:
-  
-  std::vector<int> _feed_type;  // 0 int64, 1 float
+  std::vector<std::string> _feed_name;
+  std::vector<int> _feed_type;     // 0 int64, 1 float
   std::vector<bool> _is_lod_feed;  // true lod tensor
-  std::vector<int> _capacity;  //  capacity for each tensor
-  /*
-    feed_shape_ for feeded variable
-    feed_shape_[i][j] represents the jth dim for ith input Tensor
-    if is_lod_feed_[i] == False, feed_shape_[i][0] = -1
-   */
+  std::vector<int> _capacity;      //  capacity for each tensor
+                                   /*
+                                     feed_shape_ for feeded variable
+                                     feed_shape_[i][j] represents the jth dim for ith input Tensor
+                                     if is_lod_feed_[i] == False, feed_shape_[i][0] = -1
+                                    */
   std::vector<std::vector<int>> _feed_shape;
+
+  std::vector<std::string> _fetch_name;
+  std::vector<std::vector<int>> _fetch_shape;
 };
 
 class BaseRdDict;
@@ -76,8 +79,8 @@ class Resource {
   int initialize(const std::string& path, const std::string& file);
   int cube_initialize(const std::string& path, const std::string& file);
 
-  int general_model_initialize(
-      const std::string& path, const std::string & file);
+  int general_model_initialize(const std::string& path,
+                               const std::string& file);
 
   int thread_initialize();
 
@@ -90,7 +93,7 @@ class Resource {
   std::shared_ptr<PaddleGeneralModelConfig> get_general_model_config();
 
   void print_general_model_config(
-      const std::shared_ptr<PaddleGeneralModelConfig> & config);
+      const std::shared_ptr<PaddleGeneralModelConfig>& config);
 
   std::shared_ptr<RocksDBWrapper> getDB();
 

core/predictor/src/pdserving.cpp

@@ -45,8 +45,6 @@ using baidu::paddle_serving::predictor::FLAGS_logger_path;
 using baidu::paddle_serving::predictor::FLAGS_logger_file;
 using baidu::paddle_serving::predictor::FLAGS_resource_path;
 using baidu::paddle_serving::predictor::FLAGS_resource_file;
-using baidu::paddle_serving::predictor::FLAGS_general_model_path;
-using baidu::paddle_serving::predictor::FLAGS_general_model_file;
 using baidu::paddle_serving::predictor::FLAGS_reload_interval_s;
 using baidu::paddle_serving::predictor::FLAGS_port;
 
@@ -219,11 +217,10 @@ int main(int argc, char** argv) {
 
 #ifndef BCLOUD
 
-  if (Resource::instance().general_model_initialize(
-          FLAGS_general_model_path, FLAGS_general_model_file) != 0) {
+  if (Resource::instance().general_model_initialize(FLAGS_resource_path,
+                                                    FLAGS_resource_file) != 0) {
     LOG(ERROR) << "Failed to initialize general model conf: "
-               << FLAGS_general_model_path << "/"
-               << FLAGS_general_model_file;
+               << FLAGS_resource_path << "/" << FLAGS_resource_file;
     return -1;
   }
 

examples/demo-serving/conf/general_model.prototxt

-is_lod_feed: true
-is_lod_feed: false
-is_lod_feed: true
-feed_type: 1
-feed_type: 0
-feed_type: 1
-
-feed_shape {
+feed_var {
+    name: "words"
+    is_lod_tensor: true
+    feed_type: 0
     shape: -1
 }
-
-feed_shape {
+feed_var {
+    name: "label"
+    is_lod_tensor: false
+    feed_type: 0
+    shape: 1    
+}
+fetch_var {
+    name: "cost"
     shape: 1
-    shape: 2
-    shape: 3
 }
-
-feed_shape {
-    shape: -1
+fetch_var {
+    name: "acc"
+    shape: 1
+}
+fetch_var {
+    name: "prediction"
+    shape: 2
 }

examples/demo-serving/conf/gflags.conf

 --enable_model_toolkit
 --enable_cube=false
 --enable_general_model=true
---general_model_path=./conf
---general_model_file=general_model.prototxt

examples/demo-serving/conf/resource.prototxt

 model_toolkit_path: "./conf/"
 model_toolkit_file: "model_toolkit.prototxt"
 cube_config_file: "./conf/cube.conf"
+general_model_path: "./conf/"
+general_model_file: "general_model.prototxt"
+

examples/demo-serving/conf/service.prototxt

@@ -40,6 +40,6 @@ services {
   workflows: "workflow9"
 }
 services {
-  name: "LoadGeneralModelService"
-  workflows: "workflow10"
-}
+  name: "GeneralModelService"
+  workflows: "workflow11"
+}
\ No newline at end of file

examples/demo-serving/conf/workflow.prototxt

@@ -95,7 +95,24 @@ workflows {
   name: "workflow10"
   workflow_type: "Sequence"
   nodes {
-    name: "load_general_model_conf_op"
-    type: "LoadGeneralModelConfOp"
+    name: "general_model_op"
+    type: "GeneralModelOp"
+  }
+}
+
+workflows {
+  name: "workflow11"
+  workflow_type: "Sequence"
+  nodes {
+    name: "general_reader_op"
+    type: "GeneralReaderOp"
+  }
+  nodes {
+    name: "general_infer_op"
+    type: "GeneralInferOp"
+    dependencies {
+      name: "general_reader_op"
+      mode: "RO"
+    }
   }
 }

examples/demo-serving/general_model_conf/cube.conf

-[{
-    "dict_name": "test_dict",
-    "shard": 2,
-    "dup": 1,
-    "timeout": 200,
-    "retry": 3,
-    "backup_request": 100,
-    "type": "ipport_list",
-    "load_balancer": "rr",
-    "nodes": [{
-        "ipport_list": "list://xxx.xxx.xxx.xxx:8000"
-    },{ 
-        "ipport_list": "list://xxx.xxx.xxx.xxx:8000"
-    }]  
-}]

examples/demo-serving/general_model_conf/general_model.prototxt

-is_lod_feed: true
-is_lod_feed: false
-is_lod_feed: true
-feed_type: 1
-feed_type: 0
-feed_type: 1
-
-feed_shape {
-    shape: -1
-}
-
-feed_shape {
-    shape: 1
-    shape: 2
-    shape: 3
-}
-
-feed_shape {
-    shape: -1
-}

examples/demo-serving/general_model_conf/gflags.conf

---enable_model_toolkit
---enable_cube=false
---enable_general_model=true
---general_model_path=./conf
---general_model_file=general_model.prototxt

examples/demo-serving/general_model_conf/model_toolkit.prototxt

-engines {
-  name: "general_model"
-  type: "FLUID_CPU_ANALYSIS_DIR"
-  reloadable_meta: "./data/model/paddle/fluid_time_file"
-  reloadable_type: "timestamp_ne"
-  model_data_path: "./data/model/paddle/fluid/text_classification"
-  runtime_thread_num: 0
-  batch_infer_size: 0
-  enable_batch_align: 0
-}

examples/demo-serving/general_model_conf/resource.prototxt

-model_toolkit_path: "./conf/"
-model_toolkit_file: "model_toolkit.prototxt"

examples/demo-serving/general_model_conf/service.prototxt

-port: 9292
-services {
-  name: "GeneralModelService"
-  workflows: "workflow1"
-}
\ No newline at end of file

examples/demo-serving/general_model_conf/workflow.prototxt

-workflows {
-  name: "workflow1"
-  workflow_type: "Sequence"
-  nodes {
-    name: "general_model_op"
-    type: "GeneralModelOp"
-  }
-}
\ No newline at end of file

examples/demo-serving/op/general_infer_op.cpp

+// 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.
+
+#include "examples/demo-serving/op/general_infer_op.h"
+#include <algorithm>
+#include <iostream>
+#include <memory>
+#include <sstream>
+#include "core/predictor/framework/infer.h"
+#include "core/predictor/framework/memory.h"
+#include "core/predictor/framework/resource.h"
+#include "examples/demo-serving/op/general_reader_op.h"
+
+namespace baidu {
+namespace paddle_serving {
+namespace serving {
+
+using baidu::paddle_serving::predictor::MempoolWrapper;
+using baidu::paddle_serving::predictor::general_model::Tensor;
+using baidu::paddle_serving::predictor::general_model::Response;
+using baidu::paddle_serving::predictor::general_model::FetchInst;
+using baidu::paddle_serving::predictor::InferManager;
+
+int GeneralInferOp::inference() {
+  const GeneralReaderOutput *reader_out =
+      get_depend_argument<GeneralReaderOutput>("general_reader_op");
+  if (!reader_out) {
+    LOG(ERROR) << "Failed mutable depended argument, op:"
+               << "general_reader_op";
+    return -1;
+  }
+
+  int reader_status = reader_out->reader_status;
+  if (reader_status != 0) {
+    LOG(ERROR) << "Read request wrong.";
+    return -1;
+  }
+
+  const TensorVector *in = &reader_out->tensor_vector;
+  TensorVector *out = butil::get_object<TensorVector>();
+  int batch_size = (*in)[0].shape[0];
+  // infer
+  if (InferManager::instance().infer(GENERAL_MODEL_NAME, in, out, batch_size)) {
+    LOG(ERROR) << "Failed do infer in fluid model: " << GENERAL_MODEL_NAME;
+    return -1;
+  }
+
+  Response *res = mutable_data<Response>();
+
+  for (int i = 0; i < batch_size; ++i) {
+    FetchInst *fetch_inst = res->add_insts();
+    for (int j = 0; j < out->size(); ++j) {
+      Tensor *tensor = fetch_inst->add_tensor_array();
+      tensor->set_elem_type(1);
+      if (out->at(j).lod.size() == 1) {
+        tensor->add_shape(-1);
+      } else {
+        for (int k = 1; k < out->at(j).shape.size(); ++k) {
+          tensor->add_shape(out->at(j).shape[k]);
+        }
+      }
+    }
+  }
+
+  for (int i = 0; i < out->size(); ++i) {
+    float *data_ptr = static_cast<float *>(out->at(i).data.data());
+    int cap = 1;
+    for (int j = 1; j < out->at(i).shape.size(); ++j) {
+      cap *= out->at(i).shape[j];
+    }
+    if (out->at(i).lod.size() == 1) {
+      for (int j = 0; j < batch_size; ++j) {
+        for (int k = out->at(i).lod[0][j]; k < out->at(i).lod[0][j + 1]; k++) {
+          res->mutable_insts(j)->mutable_tensor_array(i)->add_data(
+              reinterpret_cast<char *>(&(data_ptr[k])), sizeof(float));
+        }
+      }
+    } else {
+      for (int j = 0; j < batch_size; ++j) {
+        for (int k = j * cap; k < (j + 1) * cap; ++k) {
+          res->mutable_insts(j)->mutable_tensor_array(i)->add_data(
+              reinterpret_cast<char *>(&(data_ptr[k])), sizeof(float));
+        }
+      }
+    }
+  }
+  /*
+      for (size_t i = 0; i < in->size(); ++i) {
+        (*in)[i].shape.clear();
+      }
+      in->clear();
+      butil::return_object<TensorVector>(in);
+
+      for (size_t i = 0; i < out->size(); ++i) {
+        (*out)[i].shape.clear();
+      }
+      out->clear();
+      butil::return_object<TensorVector>(out);
+    }
+  */
+  return 0;
+}
+DEFINE_OP(GeneralInferOp);
+
+}  // namespace serving
+}  // namespace paddle_serving
+}  // namespace baidu

examples/demo-serving/op/general_infer_op.h

+// Copyright (c) 2019 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 <vector>
+#ifdef BCLOUD
+#ifdef WITH_GPU
+#include "paddle/paddle_inference_api.h"
+#else
+#include "paddle/fluid/inference/api/paddle_inference_api.h"
+#endif
+#else
+#include "paddle_inference_api.h"  // NOLINT
+#endif
+#include "examples/demo-serving/general_model_service.pb.h"
+
+namespace baidu {
+namespace paddle_serving {
+namespace serving {
+
+static const char* GENERAL_MODEL_NAME = "general_model";
+
+class GeneralInferOp
+    : public baidu::paddle_serving::predictor::OpWithChannel<
+          baidu::paddle_serving::predictor::general_model::Response> {
+ public:
+  typedef std::vector<paddle::PaddleTensor> TensorVector;
+
+  DECLARE_OP(GeneralInferOp);
+
+  int inference();
+};
+
+}  // namespace serving
+}  // namespace paddle_serving
+}  // namespace baidu

examples/demo-serving/op/general_model_op.cpp

@@ -14,10 +14,12 @@
 
 #include "examples/demo-serving/op/general_model_op.h"
 #include <algorithm>
-#include <sstream>
 #include <iostream>
+#include <memory>
+#include <sstream>
 #include "core/predictor/framework/infer.h"
 #include "core/predictor/framework/memory.h"
+#include "core/predictor/framework/resource.h"
 
 namespace baidu {
 namespace paddle_serving {
@@ -29,10 +31,12 @@ using baidu::paddle_serving::predictor::general_model::Request;
 using baidu::paddle_serving::predictor::general_model::FeedInst;
 using baidu::paddle_serving::predictor::general_model::Response;
 using baidu::paddle_serving::predictor::general_model::FetchInst;
+using baidu::paddle_serving::predictor::PaddleGeneralModelConfig;
 
 static std::once_flag g_proto_init_flag;
 
 int GeneralModelOp::inference() {
+  // request
   const Request *req = dynamic_cast<const Request *>(get_request_message());
 
   TensorVector *in = butil::get_object<TensorVector>();
@@ -44,6 +48,7 @@ int GeneralModelOp::inference() {
   std::vector<int> elem_size;
   std::vector<int> capacity;
 
+  // infer
   if (batch_size > 0) {
     int var_num = req->insts(0).tensor_array_size();
     VLOG(3) << "var num: " << var_num;
@@ -69,9 +74,7 @@ int GeneralModelOp::inference() {
       } else {
         lod_tensor.shape.push_back(batch_size);
         capacity[i] = 1;
-        for (int k = 0;
-             k < req->insts(0).tensor_array(i).shape_size();
-             ++k) {
+        for (int k = 0; k < req->insts(0).tensor_array(i).shape_size(); ++k) {
           int dim = req->insts(0).tensor_array(i).shape(k);
           VLOG(3) << "shape for var[" << i << "]: " << dim;
           capacity[i] *= dim;
@@ -90,10 +93,9 @@ int GeneralModelOp::inference() {
     for (int i = 0; i < var_num; ++i) {
       if (in->at(i).lod.size() == 1) {
         for (int j = 0; j < batch_size; ++j) {
-          const Tensor & tensor = req->insts(j).tensor_array(i);
+          const Tensor &tensor = req->insts(j).tensor_array(i);
           int data_len = tensor.data_size();
-          VLOG(3) << "tensor size for var[" << i << "]: "
-                  << tensor.data_size();
+          VLOG(3) << "tensor size for var[" << i << "]: " << tensor.data_size();
           int cur_len = in->at(i).lod[0].back();
           VLOG(3) << "current len: " << cur_len;
           in->at(i).lod[0].push_back(cur_len + data_len);
@@ -101,18 +103,18 @@ int GeneralModelOp::inference() {
         }
         in->at(i).data.Resize(in->at(i).lod[0].back() * elem_size[i]);
         in->at(i).shape = {in->at(i).lod[0].back(), 1};
-        VLOG(3) << "var[" << i << "] is lod_tensor and len="
-                << in->at(i).lod[0].back();
+        VLOG(3) << "var[" << i
+                << "] is lod_tensor and len=" << in->at(i).lod[0].back();
       } else {
         in->at(i).data.Resize(batch_size * capacity[i] * elem_size[i]);
-        VLOG(3) << "var[" << i << "] is tensor and capacity="
-                << batch_size * capacity[i];
+        VLOG(3) << "var[" << i
+                << "] is tensor and capacity=" << batch_size * capacity[i];
       }
     }
 
     for (int i = 0; i < var_num; ++i) {
       if (elem_type[i] == 0) {
-        int64_t * dst_ptr = static_cast<int64_t *>(in->at(i).data.data());
+        int64_t *dst_ptr = static_cast<int64_t *>(in->at(i).data.data());
         int offset = 0;
         for (int j = 0; j < batch_size; ++j) {
           for (int k = 0; k < req->insts(j).tensor_array(i).data_size(); ++k) {
@@ -126,7 +128,7 @@ int GeneralModelOp::inference() {
           }
         }
       } else {
-        float * dst_ptr = static_cast<float *>(in->at(i).data.data());
+        float *dst_ptr = static_cast<float *>(in->at(i).data.data());
         int offset = 0;
         for (int j = 0; j < batch_size; ++j) {
           for (int k = 0; k < req->insts(j).tensor_array(i).data_size(); ++k) {
@@ -149,19 +151,30 @@ int GeneralModelOp::inference() {
       return -1;
     }
 
+    // print request
+    std::ostringstream oss;
+    int64_t *example = reinterpret_cast<int64_t *>((*in)[0].data.data());
+    for (uint32_t i = 0; i < 10; i++) {
+      oss << *(example + i) << " ";
+    }
+    VLOG(3) << "msg: " << oss.str();
+
+    // infer
     if (predictor::InferManager::instance().infer(
             GENERAL_MODEL_NAME, in, out, batch_size)) {
-      LOG(ERROR) << "Failed do infer in fluid model: "
-                  << GENERAL_MODEL_NAME;
+      LOG(ERROR) << "Failed do infer in fluid model: " << GENERAL_MODEL_NAME;
       return -1;
     }
+    // print response
+    float *example_1 = reinterpret_cast<float *>((*out)[0].data.data());
+    VLOG(3) << "result: " << *example_1;
+
+    Response *res = mutable_data<Response>();
 
-    Response * res = mutable_data<Response>();
-    
     for (int i = 0; i < batch_size; ++i) {
-      FetchInst * fetch_inst = res->add_insts();
+      FetchInst *fetch_inst = res->add_insts();
       for (int j = 0; j < out->size(); ++j) {
-        Tensor * tensor = fetch_inst->add_tensor_array();
+        Tensor *tensor = fetch_inst->add_tensor_array();
         tensor->set_elem_type(1);
         if (out->at(j).lod.size() == 1) {
           tensor->add_shape(-1);
@@ -174,30 +187,29 @@ int GeneralModelOp::inference() {
     }
 
     for (int i = 0; i < out->size(); ++i) {
-      float * data_ptr = static_cast<float *>(out->at(i).data.data());
+      float *data_ptr = static_cast<float *>(out->at(i).data.data());
       int cap = 1;
-      for (int j = 0; j < out->at(i).shape.size(); ++j) {
+      for (int j = 1; j < out->at(i).shape.size(); ++j) {
         cap *= out->at(i).shape[j];
       }
       if (out->at(i).lod.size() == 1) {
         for (int j = 0; j < batch_size; ++j) {
-          for (int k = out->at(i).lod[0][j];
-               k < out->at(i).lod[0][j + 1];
+          for (int k = out->at(i).lod[0][j]; k < out->at(i).lod[0][j + 1];
                k++) {
             res->mutable_insts(j)->mutable_tensor_array(i)->add_data(
-                (char *)(&(data_ptr[k])), sizeof(float));
+                reinterpret_cast<char *>(&(data_ptr[k])), sizeof(float));
           }
         }
       } else {
         for (int j = 0; j < batch_size; ++j) {
           for (int k = j * cap; k < (j + 1) * cap; ++k) {
             res->mutable_insts(j)->mutable_tensor_array(i)->add_data(
-                (char *)(&(data_ptr[k])), sizeof(float));
+                reinterpret_cast<char *>(&(data_ptr[k])), sizeof(float));
           }
         }
       }
     }
-    
+
     for (size_t i = 0; i < in->size(); ++i) {
       (*in)[i].shape.clear();
     }

examples/demo-serving/op/general_reader_op.cpp

+// Copyright (c) 2019 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 "examples/demo-serving/op/general_reader_op.h"
+#include <algorithm>
+#include <iostream>
+#include <memory>
+#include <sstream>
+#include "core/predictor/framework/infer.h"
+#include "core/predictor/framework/memory.h"
+
+namespace baidu {
+namespace paddle_serving {
+namespace serving {
+
+using baidu::paddle_serving::predictor::MempoolWrapper;
+using baidu::paddle_serving::predictor::general_model::Tensor;
+using baidu::paddle_serving::predictor::general_model::Request;
+using baidu::paddle_serving::predictor::general_model::FeedInst;
+using baidu::paddle_serving::predictor::PaddleGeneralModelConfig;
+
+int conf_check(const Request *req,
+               const std::shared_ptr<PaddleGeneralModelConfig> &model_config) {
+  int var_num = req->insts(0).tensor_array_size();
+  if (var_num != model_config->_feed_type.size()) {
+    LOG(ERROR) << "feed var number not match.";
+    return -1;
+  }
+  for (int i = 0; i < var_num; ++i) {
+    if (model_config->_feed_type[i] !=
+        req->insts(0).tensor_array(i).elem_type()) {
+      LOG(ERROR) << "feed type not match.";
+      return -1;
+    }
+    if (model_config->_feed_shape[i].size() ==
+        req->insts(0).tensor_array(i).shape_size()) {
+      for (int j = 0; j < model_config->_feed_shape[i].size(); ++j) {
+        req->insts(0).tensor_array(i).shape(j);
+        if (model_config->_feed_shape[i][j] !=
+            req->insts(0).tensor_array(i).shape(j)) {
+          LOG(ERROR) << "feed shape not match.";
+          return -1;
+        }
+      }
+    } else {
+      LOG(ERROR) << "feed shape not match.";
+      return -1;
+    }
+  }
+  return 0;
+}
+
+int GeneralReaderOp::inference() {
+  // reade request from client
+  const Request *req = dynamic_cast<const Request *>(get_request_message());
+
+  int batch_size = req->insts_size();
+  int input_var_num = 0;
+
+  std::vector<int64_t> elem_type;
+  std::vector<int64_t> elem_size;
+  std::vector<int64_t> capacity;
+
+  GeneralReaderOutput *res = mutable_data<GeneralReaderOutput>();
+  TensorVector *in = &res->tensor_vector;
+
+  if (!res) {
+    LOG(ERROR) << "Failed get op tls reader object output";
+  }
+  if (batch_size <= 0) {
+    res->reader_status = -1;
+    return 0;
+  }
+
+  int var_num = req->insts(0).tensor_array_size();
+  VLOG(3) << "var num: " << var_num;
+  // read config
+
+  LOG(INFO) << "start to call load general model_conf op";
+  baidu::paddle_serving::predictor::Resource &resource =
+      baidu::paddle_serving::predictor::Resource::instance();
+
+  LOG(INFO) << "get resource pointer done.";
+  std::shared_ptr<PaddleGeneralModelConfig> model_config =
+      resource.get_general_model_config();
+
+  LOG(INFO) << "print general model config done.";
+
+  // check
+  res->reader_status = conf_check(req, model_config);
+  if (res->reader_status != 0) {
+    LOG(INFO) << "model conf of server:";
+    resource.print_general_model_config(model_config);
+    return 0;
+  }
+  // package tensor
+
+  elem_type.resize(var_num);
+  elem_size.resize(var_num);
+  capacity.resize(var_num);
+  paddle::PaddleTensor lod_tensor;
+  for (int i = 0; i < var_num; ++i) {
+    elem_type[i] = req->insts(0).tensor_array(i).elem_type();
+    VLOG(3) << "var[" << i << "] has elem type: " << elem_type[i];
+    if (elem_type[i] == 0) {  // int64
+      elem_size[i] = sizeof(int64_t);
+      lod_tensor.dtype = paddle::PaddleDType::INT64;
+    } else {
+      elem_size[i] = sizeof(float);
+      lod_tensor.dtype = paddle::PaddleDType::FLOAT32;
+    }
+
+    if (req->insts(0).tensor_array(i).shape(0) == -1) {
+      lod_tensor.lod.resize(1);
+      lod_tensor.lod[0].push_back(0);
+      VLOG(3) << "var[" << i << "] is lod_tensor";
+    } else {
+      lod_tensor.shape.push_back(batch_size);
+      capacity[i] = 1;
+      for (int k = 0; k < req->insts(0).tensor_array(i).shape_size(); ++k) {
+        int dim = req->insts(0).tensor_array(i).shape(k);
+        VLOG(3) << "shape for var[" << i << "]: " << dim;
+        capacity[i] *= dim;
+        lod_tensor.shape.push_back(dim);
+      }
+      VLOG(3) << "var[" << i << "] is tensor, capacity: " << capacity[i];
+    }
+    if (i == 0) {
+      lod_tensor.name = "words";
+    } else {
+      lod_tensor.name = "label";
+    }
+    in->push_back(lod_tensor);
+  }
+
+  for (int i = 0; i < var_num; ++i) {
+    if (in->at(i).lod.size() == 1) {
+      for (int j = 0; j < batch_size; ++j) {
+        const Tensor &tensor = req->insts(j).tensor_array(i);
+        int data_len = tensor.data_size();
+        VLOG(3) << "tensor size for var[" << i << "]: " << tensor.data_size();
+        int cur_len = in->at(i).lod[0].back();
+        VLOG(3) << "current len: " << cur_len;
+        in->at(i).lod[0].push_back(cur_len + data_len);
+        VLOG(3) << "new len: " << cur_len + data_len;
+      }
+      in->at(i).data.Resize(in->at(i).lod[0].back() * elem_size[i]);
+      in->at(i).shape = {in->at(i).lod[0].back(), 1};
+      VLOG(3) << "var[" << i
+              << "] is lod_tensor and len=" << in->at(i).lod[0].back();
+    } else {
+      in->at(i).data.Resize(batch_size * capacity[i] * elem_size[i]);
+      VLOG(3) << "var[" << i
+              << "] is tensor and capacity=" << batch_size * capacity[i];
+    }
+  }
+
+  for (int i = 0; i < var_num; ++i) {
+    if (elem_type[i] == 0) {
+      int64_t *dst_ptr = static_cast<int64_t *>(in->at(i).data.data());
+      int offset = 0;
+      for (int j = 0; j < batch_size; ++j) {
+        for (int k = 0; k < req->insts(j).tensor_array(i).data_size(); ++k) {
+          dst_ptr[offset + k] =
+              *(const int64_t *)req->insts(j).tensor_array(i).data(k).c_str();
+        }
+        if (in->at(i).lod.size() == 1) {
+          offset = in->at(i).lod[0][j + 1];
+        } else {
+          offset += capacity[i];
+        }
+      }
+    } else {
+      float *dst_ptr = static_cast<float *>(in->at(i).data.data());
+      int offset = 0;
+      for (int j = 0; j < batch_size; ++j) {
+        for (int k = 0; k < req->insts(j).tensor_array(i).data_size(); ++k) {
+          dst_ptr[offset + k] =
+              *(const float *)req->insts(j).tensor_array(i).data(k).c_str();
+        }
+        if (in->at(i).lod.size() == 1) {
+          offset = in->at(i).lod[0][j + 1];
+        } else {
+          offset += capacity[i];
+        }
+      }
+    }
+  }
+
+  VLOG(3) << "read data from client success";
+  // print request
+  std::ostringstream oss;
+  int64_t *example = reinterpret_cast<int64_t *>((*in)[0].data.data());
+  for (int i = 0; i < 10; i++) {
+    oss << *(example + i) << " ";
+  }
+  VLOG(3) << "head element of first feed var : " << oss.str();
+  //
+  return 0;
+}
+DEFINE_OP(GeneralReaderOp);
+}  // namespace serving
+}  // namespace paddle_serving
+}  // namespace baidu

examples/demo-serving/op/general_reader_op.h

+// Copyright (c) 2019 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 <vector>
+#ifdef BCLOUD
+#ifdef WITH_GPU
+#include "paddle/paddle_inference_api.h"
+#else
+#include "paddle/fluid/inference/api/paddle_inference_api.h"
+#endif
+#else
+#include "paddle_inference_api.h"  // NOLINT
+#endif
+#include <string>
+#include "core/predictor/framework/resource.h"
+#include "examples/demo-serving/general_model_service.pb.h"
+#include "examples/demo-serving/load_general_model_service.pb.h"
+
+namespace baidu {
+namespace paddle_serving {
+namespace serving {
+
+struct GeneralReaderOutput {
+  std::vector<paddle::PaddleTensor> tensor_vector;
+  int reader_status = 0;
+
+  void Clear() {
+    size_t tensor_count = tensor_vector.size();
+    for (size_t ti = 0; ti < tensor_count; ++ti) {
+      tensor_vector[ti].shape.clear();
+    }
+    tensor_vector.clear();
+  }
+  std::string ShortDebugString() const { return "Not implemented!"; }
+};
+
+class GeneralReaderOp : public baidu::paddle_serving::predictor::OpWithChannel<
+                            GeneralReaderOutput> {
+ public:
+  typedef std::vector<paddle::PaddleTensor> TensorVector;
+
+  DECLARE_OP(GeneralReaderOp);
+
+  int inference();
+};
+
+}  // namespace serving
+}  // namespace paddle_serving
+}  // namespace baidu

python/examples/imdb/test_client_batch.py

@@ -24,13 +24,13 @@ def batch_predict(batch_size=4):
     client.load_client_config(conf_file)
     client.connect(["127.0.0.1:8010"])
     start = time.time()
+    fetch = ["acc", "cost", "prediction"]
     feed_batch = []
     for line in sys.stdin:
         group = line.strip().split()
         words = [int(x) for x in group[1:int(group[0])]]
         label = [int(group[-1])]
         feed = {"words": words, "label": label}
-        fetch = ["acc", "cost", "prediction"]
         feed_batch.append(feed)
         if len(feed_batch) == batch_size:
             fetch_batch = client.batch_predict(
@@ -39,6 +39,11 @@ def batch_predict(batch_size=4):
                 print("{} {}".format(fetch_batch[i]["prediction"][1],
                                      feed_batch[i]["label"][0]))
             feed_batch = []
+    if len(feed_batch) > 0:
+        fetch_batch = client.batch_predict(feed_batch=feed_batch, fetch=fetch)
+        for i in range(len(feed_batch)):
+            print("{} {}".format(fetch_batch[i]["prediction"][1], feed_batch[i][
+                "label"][0]))
     cost = time.time() - start
     print("total cost : {}".format(cost))