Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix_crop

Author: wanghaoshuang

paddle/framework/backward.cc

@@ -522,7 +522,7 @@ ParamGradInfoMap AppendBackward(
       new OpDescBind("fill_constant", {}, {{"Out", {fill_one_op_out}}},
                      {{"shape", std::vector<int>{1}},
                       {"value", static_cast<float>(1.0)},
-                      {"data_type", target.GetDataType()}}));
+                      {"dtype", target.GetDataType()}}));
   // infer var type of fill_one_op
   fill_one_op->InferVarType(root_block);
 

paddle/framework/executor.cc

@@ -120,7 +120,7 @@ void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id,
 
   for (auto& op_desc : block.AllOps()) {
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
-    VLOG(10) << op->DebugString();
+    VLOG(3) << op->DebugString();
     op->Run(*local_scope, *device);
   }
   if (create_local_scope) {

paddle/framework/prune.cc

@@ -26,6 +26,8 @@ namespace framework {
 
 const std::string kFeedOpType = "feed";
 const std::string kFetchOpType = "fetch";
+const std::string kDropOutOpType = "dropout";
+const std::string kBatchNormOpType = "batch_norm";
 
 bool HasDependentVar(const OpDesc& op_desc,
                      const std::set<std::string>& dependent_vars) {
@@ -106,5 +108,26 @@ void Prune(const ProgramDesc& input, ProgramDesc* output) {
   prune_impl(input, output, 0);
 }
 
+void inference_optimize_impl(const ProgramDesc& input, ProgramDesc* output,
+                             int block_id) {
+  *output = input;
+  auto* op_field = output->mutable_blocks(block_id)->mutable_ops();
+  for (auto& op_desc : *op_field) {
+    if (op_desc.type() == kDropOutOpType ||
+        op_desc.type() == kBatchNormOpType) {
+      for (auto& attr : *op_desc.mutable_attrs()) {
+        if (attr.name() == "is_test") {
+          attr.set_b(true);
+          break;
+        }
+      }
+    }
+  }
+}
+
+void InferenceOptimize(const ProgramDesc& input, ProgramDesc* output) {
+  inference_optimize_impl(input, output, 0);
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/prune.h

@@ -22,5 +22,7 @@ namespace framework {
 
 void Prune(const ProgramDesc& input, ProgramDesc* output);
 
+void InferenceOptimize(const ProgramDesc& input, ProgramDesc* output);
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/tensor_array.cc

@@ -302,7 +302,7 @@ LoDTensor TensorArray::Stack() const {
 
   const auto& first_dims = values_.front().dims();
   // check all the values have the same shape
-  // TODO(superjom) check the same dtypes
+  // TODO(superjom) check the same data_type
   for (size_t idx = 1; idx < size(); idx++) {
     const auto& value_dims = values_[idx].dims();
     PADDLE_ENFORCE_EQ(first_dims, value_dims);

paddle/memory/CMakeLists.txt

@@ -1,6 +1,6 @@
 add_subdirectory(detail)
 
-cc_library(memory SRCS memory.cc DEPS place)
+cc_library(memory SRCS memory.cc DEPS place enforce)
 cc_library(memcpy SRCS memcpy.cc)
 
 cc_library(paddle_memory

paddle/operators/beam_search_decode_op.cc

@@ -17,6 +17,36 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+struct BeamSearchDecodeFunctor {
+  BeamSearchDecodeFunctor(const LoDTensorArray& step_ids,
+                          const LoDTensorArray& step_scores,
+                          LoDTensor* id_tensor, LoDTensor* score_tensor)
+      : step_ids_(step_ids),
+        step_scores_(step_scores),
+        id_tensor_(id_tensor),
+        score_tensor_(score_tensor) {}
+
+  template <typename T>
+  void operator()() const;
+
+  const LoDTensorArray& step_ids_;
+  const LoDTensorArray& step_scores_;
+  LoDTensor* id_tensor_;
+  LoDTensor* score_tensor_;
+};
+
+template <typename T>
+void BeamSearchDecodeFunctor::operator()() const {
+  BeamSearchDecoder<T> beam_search_decoder;
+  beam_search_decoder.PackAllSteps(step_ids_, step_scores_, id_tensor_,
+                                   score_tensor_);
+}
+
+template <>
+void BeamSearchDecodeFunctor::operator()<bool>() const {
+  PADDLE_THROW("beam search decode op does not support bool!");
+}
+
 class BeamSearchDecodeOp : public framework::OperatorBase {
  public:
   BeamSearchDecodeOp(const std::string& type,
@@ -45,9 +75,9 @@ class BeamSearchDecodeOp : public framework::OperatorBase {
     LoDTensor* sentenceIds = ctx.Output<LoDTensor>("SentenceIds");
     LoDTensor* sentenceScores = ctx.Output<LoDTensor>("SentenceScores");
 
-    BeamSearchDecoder<float> beam_search_decoder;
-    beam_search_decoder.PackAllSteps(*ids, *scores, sentenceIds,
-                                     sentenceScores);
+    framework::VisitDataType(
+        framework::ToDataType(scores->at(0).type()),
+        BeamSearchDecodeFunctor(*ids, *scores, sentenceIds, sentenceScores));
   }
 };
 

paddle/operators/bilinear_tensor_product_op.cc

@@ -77,11 +77,19 @@ class BilinearTensorProductOpMaker : public framework::OpProtoAndCheckerMaker {
     AddOutput("Out", "The output of bilinear_tensor_product operator.");
     AddComment(R"DOC(
 Bilinear Tensor Product operator.
-Given input X and Y, a 3D tensor weight, and bias. Each column of the
-output is computed by one slice i = 1, . . . , k of the tensor:
-
-    M =  (X W_i) \cdot Y
-    Out_i = \sum_i {M_i} + Bias_i
+Given input X and Y, a 3D tensor Weight and a Bias. Each column of the
+Output is computed by one slice $i = 1, . . . , k$ of the tensor:
+
+$$
+M =  (X W_i) * Y \\
+Out_i = \sum_j {M_j} + Bias_i
+$$
+
+Where $W_i$ is the $i$-th slice of Input(Weight);
+      $M_j$ is the $j$-th column of $M$;
+      $Out_i$ is the $i$-th column of Output(Out);
+      $Bias_i$ is a column vector, each element of it is equal to
+        the $i$-th element of $Bias$;
 
 )DOC");
   }

paddle/operators/cast_op.cc

@@ -25,8 +25,8 @@ class CastOpProtoMaker : public framework::OpProtoAndCheckerMaker {
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input tensor of cast op");
     AddOutput("Out", "The output tensor of cast op");
-    AddAttr<int>("out_data_type", "output data type");
-    AddAttr<int>("in_data_type", "input data type");
+    AddAttr<int>("out_dtype", "output data type");
+    AddAttr<int>("in_dtype", "input data type");
     AddComment(R"DOC(
 Cast Operator.
 
@@ -58,8 +58,8 @@ class CastOpGradMaker : public framework::SingleGradOpDescMaker {
     grad->SetType("cast");
     grad->SetInput("X", OutputGrad("Out"));
     grad->SetOutput("Out", InputGrad("X"));
-    grad->SetAttr("out_data_type", GetAttr("in_data_type"));
-    grad->SetAttr("in_data_type", GetAttr("out_data_type"));
+    grad->SetAttr("out_dtype", GetAttr("in_dtype"));
+    grad->SetAttr("in_dtype", GetAttr("out_dtype"));
     return std::unique_ptr<framework::OpDescBind>(grad);
   }
 };

paddle/operators/cast_op.h

@@ -55,7 +55,7 @@ class CastOpKernel : public framework::OpKernel<InT> {
     auto* in = context.Input<framework::Tensor>("X");
     auto* out = context.Output<framework::Tensor>("Out");
     framework::VisitDataType(
-        static_cast<framework::DataType>(context.Attr<int>("out_data_type")),
+        static_cast<framework::DataType>(context.Attr<int>("out_dtype")),
         CastOpFunctor<Place, InT>(in, out, context.device_context()));
   }
 };