add general server c++ code && fix batch predict bug

core/configure/proto/general_model_config.proto

@@ -15,12 +15,19 @@
 syntax = "proto2";
 package baidu.paddle_serving.configure;
 
-message Shape {
-  repeated int32 shape = 1;
-};
+message Shape { repeated int32 shape = 1; };
 
+message FeedVar {
+  required string name = 1;
+  required bool is_lod_tensor = 2;
+  required int32 feed_type = 3;
+  required Shape feed_shape = 4;
+}
+message FetchVar {
+  required string name = 1;
+  required Shape fetch_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;
+  required int32 rpc_timeout_ms = 2;
+  required int32 connect_retry_count = 3;
+  required int32 max_connection_per_host = 4;
+  required int32 hedge_request_timeout_ms = 5;
+  required int32 hedge_fetch_retry_count = 6;
+  required string connection_type = 7;
 };
 
 message NamingConf {
-  optional string cluster_filter_strategy = 1 [default="Default"];
-  optional string load_balance_strategy = 2 [default = "la"];
+  optional string cluster_filter_strategy = 1;
+  optional string load_balance_strategy = 2;
   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;
+  required int32 package_size = 2;
+  required string protocol = 3;
+  required int32 max_channel_per_request = 4;
 };
 
 message SplitConf {
@@ -53,12 +53,12 @@ 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; };
 
 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;
+  required string service_name = 2;
+  required string endpoint_router = 3;
   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

@@ -47,15 +47,22 @@ 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 +78,7 @@ 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 +86,8 @@ 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 +153,71 @@ 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).feed_shape().shape_size();
+           ++j) {
+        int32_t dim = model_config.feed_var(i).feed_shape().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).fetch_shape().shape_size();
+         ++j) {
+      int dim = model_config.fetch_var(i).fetch_shape().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 {
-    shape: -1
+feed_var {
+    name: "words"
+    is_lod_tensor: false
+    feed_type: 0
+    feed_shape {
+        shape: -1
+    }
 }
-
-feed_shape {
-    shape: 1
-    shape: 2
-    shape: 3
+feed_var {
+    name: "label"
+    is_lod_tensor: false
+    feed_type: 0
+    feed_shape {
+        shape: 1    
+    }
 }
-
-feed_shape {
-    shape: -1
+fetch_var {
+    name: "mean_0.tmp_0"
+    fetch_shape {
+        shape: 1
+    }
+}
+fetch_var {
+    name: "accuracy_0.tmp_0"
+    fetch_shape {
+        shape: 1
+    }
+}
+fetch_var {
+    name: "fc_1.tmp_2"
+    fetch_shape {
+        shape: 1
+    }
 }

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"
-}
\ No newline at end of file
+  name: "GeneralModelService"
+  workflows: "workflow11"
+}

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/op/general_model_op.cpp

@@ -15,9 +15,11 @@
 #include "examples/demo-serving/op/general_model_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"
 
 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 +48,17 @@ int GeneralModelOp::inference() {
   std::vector<int> elem_size;
   std::vector<int> capacity;
 
+  // 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) << "get general model config pointer done.";
+  resource.print_general_model_config(model_config);
+
+  // infer
   if (batch_size > 0) {
     int var_num = req->insts(0).tensor_array_size();
     VLOG(3) << "var num: " << var_num;
@@ -146,11 +161,23 @@ 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;
       return -1;
     }
+    // print response
+    float *example_1 = reinterpret_cast<float *>((*out)[0].data.data());
+    VLOG(3) << "result: " << *example_1;
 
     Response *res = mutable_data<Response>();
 

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))