Merge branch 'develop' into python_map

Author: wanghaox

CMakeLists.txt

@@ -60,7 +60,7 @@ option(USE_NNPACK       "Compile PaddlePaddle with NNPACK library"      OFF)
 option(WITH_DISTRIBUTE  "Compile with grpc distributed support"         OFF)
 option(USE_EIGEN_FOR_BLAS   "Use matrix multiplication in Eigen"        OFF)
 option(WITH_ARM_FP16    "Use half precision support on armv8.2-a cpu"   OFF)
-option(WITH_FAST_BUNDLE_TEST    "Bundle tests that can be run in a single process together to reduce launch overhead"   ON)
+option(WITH_FAST_BUNDLE_TEST    "Bundle tests that can be run in a single process together to reduce launch overhead"   OFF)
 
 # CMAKE_BUILD_TYPE
 if(NOT CMAKE_BUILD_TYPE)

doc/howto/capi/workflow_of_capi_cn.md

@@ -65,6 +65,7 @@
     output_file = "output.paddle.model"
     merge_v2_model(net, param_file, output_file)
     ```
+
     对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)这个示例，可直接运行 `python` [merge_v2_model.py](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense/merge_v2_model.py)。序列化结果会写入当前运行目录下的`output.paddle.model`文件中。使用这种方式，运行时C-API可以通过指定`output.paddle.model`文件的路径来加载预测模型。
 
 #### 注意事项

paddle/fluid/framework/executor.cc

@@ -58,13 +58,13 @@ static void CreateTensor(Variable* var, proto::VarType::Type var_type) {
     var->GetMutable<ReaderHolder>();
   } else if (var_type == proto::VarType::CHANNEL) {
     var->GetMutable<ChannelHolder>();
-  } else if (var_type == proto::VarType::NCCL_COM) {
-    // GetMutable will be called in ncclInit
+  } else if (var_type == proto::VarType::RAW) {
+    // GetMutable will be called in operator
   } else {
     PADDLE_THROW(
         "Variable type %d is not in "
         "[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, "
-        "LOD_RANK_TABLE, PLACE_LIST, READER, CHANNEL, NCCL_COM]",
+        "LOD_RANK_TABLE, PLACE_LIST, READER, CHANNEL, RAW]",
         var_type);
   }
 }

paddle/fluid/framework/framework.proto

@@ -113,7 +113,10 @@ message VarType {
     PLACE_LIST = 14;
     READER = 15;
     CHANNEL = 16;
-    NCCL_COM = 17;
+    // Any runtime decided variable type is raw
+    // raw variables should manage their own allocations
+    // in operators like nccl_op
+    RAW = 17;
   }
 
   required Type type = 1;

paddle/fluid/framework/lod_tensor.cc

@@ -31,8 +31,14 @@ std::ostream &operator<<(std::ostream &os, const LoD &lod) {
   os << "{";
   for (auto &v : lod) {
     os << "{";
+    bool is_first = true;
     for (auto &i : v) {
-      os << i << ",";
+      if (is_first) {
+        os << i;
+        is_first = false;
+      } else {
+        os << ", " << i;
+      }
     }
     os << "}";
   }

paddle/fluid/inference/io.cc

@@ -32,23 +32,11 @@ void ReadBinaryFile(const std::string& filename, std::string& contents) {
   inputfs.close();
 }
 
-bool IsParameter(const framework::VarDesc* var,
-                 const framework::ProgramDesc& main_program) {
-  if (var->Persistable()) {
-    // There are many unreachable variables in the program
-    for (size_t i = 0; i < main_program.Size(); ++i) {
-      const framework::BlockDesc& block = main_program.Block(i);
-      for (auto* op : block.AllOps()) {
-        if (op->Type() == framework::kFeedOpType) {
-          continue;
-        }
-        for (auto input_argument_name : op->InputArgumentNames()) {
-          if (input_argument_name == var->Name()) {
-            return true;
-          }
-        }
-      }
-    }
+bool IsPersistable(const framework::VarDesc* var) {
+  if (var->Persistable() &&
+      var->GetType() != framework::proto::VarType::FEED_MINIBATCH &&
+      var->GetType() != framework::proto::VarType::FETCH_LIST) {
+    return true;
   }
   return false;
 }
@@ -65,8 +53,8 @@ void LoadPersistables(framework::Executor& executor,
   std::vector<std::string> paramlist;
 
   for (auto* var : global_block.AllVars()) {
-    if (IsParameter(var, main_program)) {
-      VLOG(3) << "parameter's name: " << var->Name();
+    if (IsPersistable(var)) {
+      VLOG(3) << "persistable variable's name: " << var->Name();
 
       framework::VarDesc* new_var = load_block->Var(var->Name());
       new_var->SetShape(var->GetShape());
@@ -101,7 +89,6 @@ void LoadPersistables(framework::Executor& executor,
 
   executor.Run(*load_program, &scope, 0, true, true);
 
-  VLOG(3) << "Ran loading successfully";
   delete load_program;
 }
 

paddle/fluid/inference/tests/book/CMakeLists.txt

@@ -30,5 +30,5 @@ inference_test(label_semantic_roles)
 inference_test(recognize_digits ARGS mlp conv)
 inference_test(recommender_system)
 #inference_test(rnn_encoder_decoder)
-inference_test(understand_sentiment)
+inference_test(understand_sentiment ARGS conv)
 inference_test(word2vec)

paddle/fluid/inference/tests/book/test_inference_label_semantic_roles.cc

@@ -32,16 +32,42 @@ TEST(inference, label_semantic_roles) {
   paddle::framework::LoDTensor word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1,
       ctx_p2, mark;
   paddle::framework::LoD lod{{0, 4, 10}};
-
-  SetupLoDTensor(word, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(
-      predicate, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_n2, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_n1, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_0, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_p1, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_p2, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(mark, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
+  int64_t word_dict_len = 44068;
+  int64_t predicate_dict_len = 3162;
+  int64_t mark_dict_len = 2;
+
+  SetupLoDTensor(word,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(predicate,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(predicate_dict_len - 1));
+  SetupLoDTensor(ctx_n2,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(ctx_n1,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(ctx_0,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(ctx_p1,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(ctx_p2,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(mark,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(mark_dict_len - 1));
 
   std::vector<paddle::framework::LoDTensor*> cpu_feeds;
   cpu_feeds.push_back(&word);

paddle/fluid/inference/tests/book/test_inference_understand_sentiment.cc

@@ -31,7 +31,12 @@ TEST(inference, understand_sentiment) {
 
   paddle::framework::LoDTensor words;
   paddle::framework::LoD lod{{0, 4, 10}};
-  SetupLoDTensor(words, lod, static_cast<int64_t>(0), static_cast<int64_t>(10));
+  int64_t word_dict_len = 5147;
+
+  SetupLoDTensor(words,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
 
   std::vector<paddle::framework::LoDTensor*> cpu_feeds;
   cpu_feeds.push_back(&words);

paddle/fluid/inference/tests/book/test_inference_word2vec.cc

@@ -31,12 +31,12 @@ TEST(inference, word2vec) {
 
   paddle::framework::LoDTensor first_word, second_word, third_word, fourth_word;
   paddle::framework::LoD lod{{0, 1}};
-  int64_t dict_size = 2072;  // Hard-coding the size of dictionary
+  int64_t dict_size = 2073;  // The size of dictionary
 
-  SetupLoDTensor(first_word, lod, static_cast<int64_t>(0), dict_size);
-  SetupLoDTensor(second_word, lod, static_cast<int64_t>(0), dict_size);
-  SetupLoDTensor(third_word, lod, static_cast<int64_t>(0), dict_size);
-  SetupLoDTensor(fourth_word, lod, static_cast<int64_t>(0), dict_size);
+  SetupLoDTensor(first_word, lod, static_cast<int64_t>(0), dict_size - 1);
+  SetupLoDTensor(second_word, lod, static_cast<int64_t>(0), dict_size - 1);
+  SetupLoDTensor(third_word, lod, static_cast<int64_t>(0), dict_size - 1);
+  SetupLoDTensor(fourth_word, lod, static_cast<int64_t>(0), dict_size - 1);
 
   std::vector<paddle::framework::LoDTensor*> cpu_feeds;
   cpu_feeds.push_back(&first_word);