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

AUTHORS.md

@@ -22,6 +22,7 @@
 | jczaja | Jacek Czaja |
 | JiayiFeng | Jia-Yi Feng |
 | kbinias | Krzysztof Binias |
+| kexinzhao | Ke-Xin Zhao |
 | kuke | Yi-Bing Liu |
 | lcy-seso | Ying Cao |
 | lipeng-unisound | Peng Li |

cmake/inference_lib.cmake

@@ -155,6 +155,15 @@ copy(inference_lib DEPS paddle_fluid_shared paddle_fluid
   DSTS ${dst_dir}/${module} ${dst_dir}/${module}
 )
 
+if(WITH_CONTRIB)
+   set(contrib_dst_dir "${FLUID_INSTALL_DIR}/contrib/inference")
+   copy(contrib_inference_lib DEPS paddle_inference_api
+        SRCS ${PADDLE_SOURCE_DIR}/paddle/contrib/inference/paddle_inference_api.h
+        ${PADDLE_BINARY_DIR}/paddle/contrib/inference/libpaddle_inference_api.*
+        DSTS ${contrib_dst_dir} ${contrib_dst_dir}
+   )
+endif()
+
 set(module "platform")
 copy(platform_lib DEPS profiler_py_proto
   SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/dynload/*.h ${src_dir}/${module}/details/*.h

doc/fluid/api/layers.rst

@@ -342,6 +342,12 @@ conv2d
 ..  autofunction:: paddle.fluid.layers.conv2d
     :noindex:
 
+conv3d
+------
+
+..  autofunction:: paddle.fluid.layers.conv3d
+    :noindex:
+
 sequence_pool
 -------------
 
@@ -366,6 +372,12 @@ pool2d
 ..  autofunction:: paddle.fluid.layers.pool2d
     :noindex:
 
+pool3d
+------
+
+..  autofunction:: paddle.fluid.layers.pool3d
+    :noindex:
+
 batch_norm
 ----------
 
@@ -384,6 +396,13 @@ conv2d_transpose
 ..  autofunction:: paddle.fluid.layers.conv2d_transpose
     :noindex:
 
+conv3d_transpose
+----------------
+
+..  autofunction:: paddle.fluid.layers.conv2d_transpose
+    :noindex:
+
+
 sequence_expand
 ---------------
 

doc/v2/dev/contribute_to_paddle_cn.md

@@ -104,7 +104,7 @@ no changes added to commit (use "git add" and/or "git commit -a")
 ➜  docker run -it -v $(pwd):/paddle paddle:latest-dev bash -c "cd /paddle/build && ctest"
 ```
 
-关于构建和测试的更多信息，请参见[这篇文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_cn.rst)。
+关于构建和测试的更多信息，请参见[使用Docker安装运行](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/v2/build_and_install/docker_install_cn.rst)。
 
 ## 提交（commit）
 

paddle/contrib/tape/CMakeLists.txt

@@ -17,7 +17,7 @@ if(APPLE)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move")
 endif(APPLE)
 
-cc_library(tape_variable SRCS variable.cc DEPS ${FLUID_CORE_MODULES} device_context)
+cc_library(tape_variable SRCS variable.cc DEPS ${FLUID_CORE_MODULES} device_context framework_proto proto_desc operator)
 cc_library(tape SRCS tape.cc DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB} tape_variable)
 
 cc_test(test_tape

paddle/fluid/inference/tensorrt/convert/op_converter.h

@@ -64,7 +64,8 @@ class OpConverter {
     (*it)(op, scope, test_mode);
   }
 
-  // convert fluid block to tensorrt network
+  // Convert a fluid block to tensorrt network, NOTE it just convert operators,
+  // the INetwork's inputs and outputs should specified in some other modules.
   void ConvertBlock(const framework::proto::BlockDesc& block,
                     const std::unordered_set<std::string>& parameters,
                     const framework::Scope& scope, TensorRTEngine* engine) {

paddle/fluid/inference/tensorrt/engine.h

@@ -51,11 +51,12 @@ class TensorRTEngine : public EngineBase {
     nvinfer1::Weights w_;
   };
 
-  TensorRTEngine(int max_batch, int max_workspace, cudaStream_t* stream,
+  TensorRTEngine(int max_batch, int max_workspace,
+                 cudaStream_t* stream = nullptr,
                  nvinfer1::ILogger& logger = NaiveLogger::Global())
       : max_batch_(max_batch),
         max_workspace_(max_workspace),
-        stream_(stream),
+        stream_(stream ? stream : &default_stream_),
         logger_(logger) {}
 
   virtual ~TensorRTEngine();
@@ -121,6 +122,8 @@ class TensorRTEngine : public EngineBase {
   // the max memory size the engine uses
   int max_workspace_;
   cudaStream_t* stream_;
+  // If stream_ is not set from outside, hold its own stream.
+  cudaStream_t default_stream_;
   nvinfer1::ILogger& logger_;
 
   std::vector<Buffer> buffers_;
@@ -165,20 +168,31 @@ class TensorRTEngine : public EngineBase {
  */
 class TRT_EngineManager {
  public:
-  TensorRTEngine* Create(int max_batch, int max_workspace,
-                         cudaStream_t* stream) {
-    engines_.emplace_back(new TensorRTEngine(max_batch, max_workspace, stream));
-    return engines_.back().get();
+  bool HasEngine(const std::string& name) const {
+    return engines_.count(name) != 0;
+  }
+
+  // Get an engine called `name`.
+  TensorRTEngine* Get(const std::string& name) const {
+    return engines_.at(name).get();
+  }
+
+  // Create or get an engine called `name`
+  TensorRTEngine* Create(int max_batch, int max_workspace, cudaStream_t* stream,
+                         const std::string& name) {
+    auto* p = new TensorRTEngine(max_batch, max_workspace, stream);
+    engines_[name].reset(p);
+    return p;
   }
 
   void DeleteALl() {
-    for (auto& ptr : engines_) {
-      ptr.reset(nullptr);
+    for (auto& item : engines_) {
+      item.second.reset(nullptr);
     }
   }
 
  private:
-  std::vector<std::unique_ptr<TensorRTEngine>> engines_;
+  std::unordered_map<std::string, std::unique_ptr<TensorRTEngine>> engines_;
 };
 
 }  // namespace tensorrt

paddle/fluid/operators/activation_op.cc

@@ -252,15 +252,14 @@ class SoftShrinkOpMaker : public framework::OpProtoAndCheckerMaker {
     AddOutput("Out", "Output of Softshrink operator");
     AddAttr<float>("lambda", "non-negative offset").SetDefault(0.5f);
     AddComment(R"DOC(
-Softshrink Activation Operator.
+:strong:`Softshrink Activation Operator`
 
-$$
-out = \begin{cases} 
+..  math::
+    out = \begin{cases} 
          x - \lambda, \text{if } x > \lambda \\
          x + \lambda, \text{if } x < -\lambda \\
          0,  \text{otherwise}
          \end{cases}
-$$
 
 )DOC");
   }
@@ -271,18 +270,18 @@ class HardShrinkOpMaker : public framework::OpProtoAndCheckerMaker {
   void Make() override {
     AddInput("X", "Input of HardShrink operator");
     AddOutput("Out", "Output of HardShrink operator");
-    AddAttr<float>("threshold", "The value of threshold for HardShrink")
+    AddAttr<float>("threshold",
+                   "The value of threshold for HardShrink. [default: 0.5]")
         .SetDefault(0.5f);
     AddComment(R"DOC(
-HardShrink Activation Operator.
+:strong:`HardShrink activation operator`
 
-$$
-out = \begin{cases} 
+..  math::
+    out = \begin{cases}
             x, \text{if } x > \lambda \\
             x, \text{if } x < -\lambda \\
             0,  \text{otherwise}
           \end{cases}
-$$
 
 )DOC");
   }
@@ -394,18 +393,18 @@ class ThresholdedReluOpMaker : public framework::OpProtoAndCheckerMaker {
   void Make() override {
     AddInput("X", "Input of ThresholdedRelu operator");
     AddOutput("Out", "Output of ThresholdedRelu operator");
-    AddAttr<float>("threshold", "The threshold location of activation")
+    AddAttr<float>("threshold",
+                   "The threshold location of activation. [default 1.0].")
         .SetDefault(1.0f);
     AddComment(R"DOC(
-ThresholdedRelu Activation Operator.
+:strong:`ThresholdedRelu activation operator`
 
-$$
-out = \begin{cases} 
+..  math::
+
+    out = \begin{cases}
              x,  \text{if } x > threshold \\
              0,  \text{otherwise}
           \end{cases}
-$$
-
 )DOC");
   }
 };

paddle/fluid/operators/compare_op.cc

@@ -23,30 +23,26 @@ class CompareOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
     OpComment comment;
-    AddInput("X",
-             string::Sprintf("(LoDTensor) the left hand operand of %s operator",
+    AddInput("X", string::Sprintf("the left hand operand of %s operator",
                                   comment.type));
-    AddInput("Y", string::Sprintf(
-                      "(LoDTensor) the right hand operand of %s operator",
+    AddInput("Y", string::Sprintf("the right hand operand of %s operator",
                                   comment.type));
     AddAttr<bool>("force_cpu",
-                  "(bool, default false) Force fill output variable to cpu "
+                  "Force fill output variable to cpu "
                   "memory. Otherwise, fill output variable to the running "
-                  "device")
-        .SetDefault(false);
-    AddOutput("Out", string::Sprintf(
-                         "(LoDTensor) n-dim bool tensor. Each element is %s",
+                  "device [default true].")
+        .SetDefault(true);
+    AddOutput("Out", string::Sprintf("n-dim bool tensor. Each element is %s",
                                      comment.equation));
-    AddComment(string::Sprintf(R"DOC(%s Operator
-
+    AddComment(string::Sprintf(R"DOC(
 It operates element-wise on X and Y, and returns the Out. Each of them is a
 N-dim tensor. X and Y could be any type.  The each element of the Out tensor is
-calculated by %s
+calculated by $%s$
 )DOC",
-                               comment.type, comment.equation));
-    AddAttr<int>("axis",
-                 "(int, default -1). The start dimension index "
-                 "for broadcasting Y onto X.")
+                               comment.equation));
+    AddAttr<int>(
+        "axis",
+        "The start dimension index for broadcasting Y onto X. [default -1]")
         .SetDefault(-1)
         .EqualGreaterThan(-1);
   }

paddle/fluid/operators/concat_op.cc

@@ -107,7 +107,13 @@ REGISTER_OPERATOR(concat, ops::ConcatOp, ops::ConcatOpMaker,
                       false> /* set false to disable empty grad */);
 REGISTER_OPERATOR(concat_grad, ops::ConcatOpGrad);
 REGISTER_OP_CPU_KERNEL(
-    concat, ops::ConcatKernel<paddle::platform::CPUDeviceContext, float>);
+    concat, ops::ConcatKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::ConcatKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::ConcatKernel<paddle::platform::CPUDeviceContext, int64_t>,
+    ops::ConcatKernel<paddle::platform::CPUDeviceContext, int>);
 REGISTER_OP_CPU_KERNEL(
     concat_grad,
-    ops::ConcatGradKernel<paddle::platform::CPUDeviceContext, float>);
+    ops::ConcatGradKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::ConcatGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::ConcatGradKernel<paddle::platform::CPUDeviceContext, int64_t>,
+    ops::ConcatGradKernel<paddle::platform::CPUDeviceContext, int>);

paddle/fluid/operators/concat_op.cu.cc

@@ -15,7 +15,13 @@ limitations under the License. */
 #include "paddle/fluid/operators/concat_op.h"
 namespace ops = paddle::operators;
 REGISTER_OP_CUDA_KERNEL(
-    concat, ops::ConcatKernel<paddle::platform::CUDADeviceContext, float>);
+    concat, ops::ConcatKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ConcatKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ConcatKernel<paddle::platform::CUDADeviceContext, int64_t>,
+    ops::ConcatKernel<paddle::platform::CUDADeviceContext, int>);
 REGISTER_OP_CUDA_KERNEL(
     concat_grad,
-    ops::ConcatGradKernel<paddle::platform::CUDADeviceContext, float>);
+    ops::ConcatGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ConcatGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ConcatGradKernel<paddle::platform::CUDADeviceContext, int64_t>,
+    ops::ConcatGradKernel<paddle::platform::CUDADeviceContext, int>);

paddle/fluid/operators/cumsum_op.cc

@@ -30,19 +30,19 @@ class CumOp : public framework::OperatorWithKernel {
 class CumsumOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
-    AddInput("X", "Input of Cumsum operator");
-    AddOutput("Out", "Output of Cumsum operator");
+    AddInput("X", "Input of cumsum operator");
+    AddOutput("Out", "Output of cumsum operator");
     AddAttr<int>("axis",
-                 "(int, default -1). The dimenstion to accumulate along. "
-                 "-1 means the last dimenstion")
+                 "The dimenstion to accumulate along. -1 means the last "
+                 "dimenstion [default -1].")
         .SetDefault(-1)
         .EqualGreaterThan(-1);
     AddAttr<bool>("exclusive",
-                  "bool, default false). Whether to perform exclusive cumsum")
+                  "Whether to perform exclusive cumsum. [default false].")
         .SetDefault(false);
     AddAttr<bool>("reverse",
-                  "bool, default false). If true, the cumsum is performed in "
-                  "the reversed direction")
+                  "If true, the cumsum is performed in the reversed direction. "
+                  "[default false].")
         .SetDefault(false);
     AddComment(R"DOC(
 The cumulative sum of the elements along a given axis.

paddle/fluid/operators/get_places_op.cc

@@ -85,7 +85,7 @@ class GetPlacesOpProtoMaker : public framework::OpProtoAndCheckerMaker {
         .InEnum({"CUDA", "CPU", "AUTO"})
         .SetDefault("AUTO");
     AddComment(R"DOC(
-Returns a list of places based on flags. The list will be used for parallel
+Returns a list of places based on arguments. The list will be used for parallel
 execution.
 )DOC");
   }

paddle/fluid/operators/layer_norm_op.cc

@@ -62,36 +62,33 @@ class LayerNormOp : public framework::OperatorWithKernel {
 class LayerNormOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
-    AddInput("X", "(LoDTensor) The input tensor.");
+    AddInput("X", "The input tensor.");
     AddInput("Scale",
-             "(Tensor, optional) Scale is a 1-dimensional tensor of size "
+             "(optional) Scale is a 1-dimensional tensor of size "
              "H(`begin_norm_axis` splits the tensor(`X`) to a matrix [N,H])."
              "It is applied to the output.")
         .AsDispensable();
     AddInput("Bias",
-             "(Tensor, optional) Bias is a 1-dimensional tensor of size "
+             "(optional) Bias is a 1-dimensional tensor of size "
              "H(`begin_norm_axis` splits the tensor(`X`) to a matrix [N,H])."
              "It is applied to the output.")
         .AsDispensable();
-    AddOutput("Y", "(LoDTensor) Result after normalization.");
-    AddOutput("Mean", "(Tensor) Mean of the current mini batch.")
-        .AsIntermediate();
-    AddOutput("Variance", "(Tensor) Variance of the current mini batch.")
+    AddOutput("Y", "Result after normalization.");
+    AddOutput("Mean", "Mean of the current mini batch.").AsIntermediate();
+    AddOutput("Variance", "Variance of the current mini batch.")
         .AsIntermediate();
 
     AddAttr<float>("epsilon",
-                   "(float, default 1e-5) Constant for "
-                   "numerical stability")
+                   "Constant for numerical stability [default 1e-5].")
         .SetDefault(1e-5)
         .AddCustomChecker([](const float &epsilon) {
           PADDLE_ENFORCE(epsilon >= 0.0f && epsilon <= 0.001f,
                          "'epsilon' should be between 0.0 and 0.001.");
         });
     AddAttr<int>("begin_norm_axis",
-                 "(int default:1), the "
-                 "axis of `begin_norm_axis ... Rank(X) - 1` will be "
+                 "the axis of `begin_norm_axis ... Rank(X) - 1` will be "
                  "normalized. `begin_norm_axis` splits the tensor(`X`) to a "
-                 "matrix [N,H].")
+                 "matrix [N,H]. [default 1].")
         .SetDefault(1)
         .AddCustomChecker([](const int &begin_norm_axis) {
           PADDLE_ENFORCE_GT(begin_norm_axis, 0,
@@ -99,10 +96,14 @@ class LayerNormOpMaker : public framework::OpProtoAndCheckerMaker {
         });
 
     AddComment(R"DOC(
-Layer Normalization.
-Layer Norm has been implemented as discussed in the paper:
-https://arxiv.org/abs/1607.06450
-...
+Assume feature vectors exist on dimensions
+:attr:`begin_norm_axis ... rank(input)` and calculate the moment statistics
+along these dimensions for each feature vector :math:`a` with size
+:math:`H`, then normalize each feature vector using the corresponding
+statistics. After that, apply learnable gain and bias on the normalized
+tensor to scale and shift if :attr:`scale` and :attr:`shift` are set.
+
+Refer to `Layer Normalization <https://arxiv.org/pdf/1607.06450v1.pdf>`_
 )DOC");
   }
 };

paddle/fluid/operators/mean_op.cc

@@ -33,12 +33,10 @@ class MeanOp : public framework::OperatorWithKernel {
 class MeanOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
-    AddInput("X", "The input of mean op");
-    AddOutput("Out", "The output of mean op").Reuse("X");
+    AddInput("X", "(Tensor) The input of mean op");
+    AddOutput("Out", "(Tensor) The output of mean op").Reuse("X");
     AddComment(R"DOC(
-Mean Operator.
-
-Out is a scalar which is the mean of all elements in X. 
+Mean Operator calculates the mean of all elements in X.
 
 )DOC");
   }

paddle/fluid/operators/multiplex_op.cc

@@ -62,26 +62,46 @@ class MultiplexOp : public framework::OperatorWithKernel {
 class MultiplexOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
-    AddInput("Ids", "The index tensor of multiplex operator.");
-    AddInput("X", "The candidate tensors of multiplex operator.")
+    AddInput("Ids",
+             "Tensor<int32>, index variable which is a 2-D tensor with shape "
+             "[M, 1] where M is the batch size.");
+    AddInput("X",
+             "A list of variables to gather from. All variables have the same "
+             "shape and the rank is at least 2.")
         .AsDuplicable();
     AddOutput("Out", "The output tensor of multiplex operator.");
     AddComment(R"DOC(
-Multiplex Operator.
-
-Multiplex multiple tensors according to the index provided by the index tensor.
-
-Ids: the index tensor.
-X[0 : N - 1]: the candidate tensors for output (N >= 2).
-For each index i from 0 to batchSize - 1, the output is the i-th row of the
+Referring to the given index variable, this layer selects rows from the
+input variables to construct a multiplex variable. Assuming that there are
+:math:`m` input variables and :math:`I_i` represents the i-th input
+variable and :math:`i` is in [0, :math:`m`). All input variables are
+tensors with same shape [:math:`d_0`, :math:`d_1`, ..., :math:`d_R`].
+Please note that rank of the input tensor should be at least 2. Each input
+variable will be treated as a 2-D matrix with shape [:math:`M`, :math:`N`]
+where :math:`M` for :math:`d_0` and :math:`N` for :math:`d_1` * :math:`d_2`
+* ... * :math:`d_R`. Let :math:`I_i[j]` be the j-th row of the i-th input
+variable. The given index variable should be a 2-D tensor with shape
+[:math:`M`, 1]. Let `ID[i]` be the i-th index value of the index variable.
+Then the output variable will be a tensor with shape [:math:`d_0`,
+:math:`d_1`, ..., :math:`d_R`]. If we treat the output tensor as a 2-D
+matrix with shape [:math:`M`, :math:`N`] and let :math:`O[i]` be the i-th
+row of the matrix, then `O[i]` is equal to :math:`I_{ID[i]}[i]`.
+
+* Ids: the index tensor.
+
+* X[0 : N - 1]: the candidate tensors for output (N >= 2).
+
+* For each index i from 0 to batchSize - 1, the output is the i-th row of the
 the (Ids[i])-th tensor.
 
 For i-th row of the output tensor:
 
-$$y[i] = x_{k}[i]$$
+$$
+y[i] = x_{k}[i]
+$$
 
-where `y` is the output tensor, `x_{k}` is the k-th input tensor,
-and `k = Ids[i]`.
+where $y$ is the output tensor, $x_{k}$ is the k-th input tensor,
+and $k = Ids[i]$.
 
 )DOC");
   }

paddle/fluid/operators/reader/create_recordio_file_reader_op.cc

@@ -78,11 +78,15 @@ class CreateRecordIOReaderOp : public framework::OperatorBase {
 class CreateRecordIOReaderOpMaker : public FileReaderMakerBase {
  protected:
   void Apply() override {
-    AddAttr<std::string>("filename", "The filename of record io reader");
+    AddAttr<std::string>(
+        "filename",
+        "The filename of record file. This file will given to reader.");
     AddComment(R"DOC(
-      CreateRecordIOReader Operator
+Open a recordio file and return the reader object. The returned reader object
+is thread-safe.
 
-      Create a reader from a record io file
+NOTE: This is a very low-level API. It is used for debugging data file or
+training. Please use `open_files` instead of this API for production usage.
     )DOC");
   }
 };

paddle/fluid/operators/reader/reader_op_registry.cc

@@ -54,7 +54,7 @@ std::unique_ptr<framework::ReaderBase> CreateReaderByFileName(
 }
 
 void FileReaderMakerBase::Make() {
-  AddOutput("Out", "(ReaderHolder) The created random reader.").AsDuplicable();
+  AddOutput("Out", "(ReaderHolder): The created random reader.").AsDuplicable();
   AddAttr<std::vector<int>>("shape_concat", "The concat of all data's shapes.");
   AddAttr<std::vector<int>>(
       "ranks",

paddle/fluid/operators/row_conv_op.cc

@@ -78,23 +78,23 @@ class RowConvOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
     AddInput("X",
-             "(LoDTensor), the input(X) is a LodTensor, which supports "
+             "the input(X) is a LodTensor, which supports "
              "variable time-length input sequences. The underlying tensor "
              "in this LoDTensor is a matrix with shape (T x N), where T "
              "is the total time steps in this mini-batch and N is the input "
              "data dimension.");
     AddInput("Filter",
-             "(Tensor), the input(Filter) is a learnable parameter. It "
+             "the input(Filter) is a learnable parameter. It "
              "is a 2-D tensor with shape (future_context x N), where, "
              "future_context is the future context length and N is the data "
              "dimension.");
     AddOutput("Out",
-              "(LoDTensor), the output(Out) is a LodTensor, which supports "
+              "the output(Out) is a LodTensor, which supports "
               "variable time-length input sequences. The underlying tensor "
               "in this LodTensor is a matrix with shape T x N, i.e., the "
               "same shape as X.");
     AddComment(R"DOC(
-Row-convolution Operator.
+:strong:`Row-convolution operator`
 
 The row convolution is called lookahead convolution.  This operator was 
 introduced in the following paper for DeepSpeech2:
@@ -114,9 +114,23 @@ and a filter ($W$) of size $context \times d$,
 the output sequence is convolved as:
 
 $$
-out_{i, :} = \sum_{j=i}^{i + context} in_{j,:} \dot W_{i-j, :}
+out_{i, :} = \\sum_{j=i}^{i + context} in_{j,:} \\cdot W_{i-j, :}
 $$
 
+In the above equation:
+
+* $Out_{i}$: The i-th row of output variable with shape [1, D].
+
+* $\\tau$: Future context size.
+
+* $X_{j}$: The j-th row of input variable with shape [1, D].
+
+* $W_{i-j}$: The (i-j)-th row of parameters with shape [1, D].
+
+More details about row_conv please refer to
+the design document
+https://github.com/PaddlePaddle/Paddle/issues/2228#issuecomment-303903645 .
+
 )DOC");
   }
 };

paddle/fluid/operators/split_op.cc

@@ -115,4 +115,7 @@ USE_CPU_ONLY_OP(concat);
 
 REGISTER_OPERATOR(split, ops::SplitOp, ops::SplitOpMaker, ops::SplitGradMaker);
 REGISTER_OP_CPU_KERNEL(split,
-                       ops::SplitOpKernel<paddle::platform::CPUPlace, float>);
+                       ops::SplitOpKernel<paddle::platform::CPUPlace, double>,
+                       ops::SplitOpKernel<paddle::platform::CPUPlace, float>,
+                       ops::SplitOpKernel<paddle::platform::CPUPlace, int64_t>,
+                       ops::SplitOpKernel<paddle::platform::CPUPlace, int>);

paddle/fluid/operators/split_op.cu.cc

@@ -15,4 +15,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/split_op.h"
 namespace ops = paddle::operators;
 REGISTER_OP_CUDA_KERNEL(
-    split, ops::SplitOpKernel<paddle::platform::CUDADeviceContext, float>);
+    split, ops::SplitOpKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::SplitOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SplitOpKernel<paddle::platform::CUDADeviceContext, int64_t>,
+    ops::SplitOpKernel<paddle::platform::CUDADeviceContext, int>);

paddle/fluid/operators/tensorrt_engine_op.cc

@@ -66,17 +66,25 @@ nvinfer1::Dims Vec2TRT_Dims(const std::vector<int64_t> &shape) {
 }  // namespace
 
 template <typename DeviceContext, typename T>
-void paddle::operators::TensorRTEngineKernel<DeviceContext, T>::Prepare(
+void TensorRTEngineKernel<DeviceContext, T>::Prepare(
     const framework::ExecutionContext &context) const {
   VLOG(4) << "Prepare engine";
   // Get the ProgramDesc and pass to convert.
   framework::proto::BlockDesc block_desc;
   block_desc.ParseFromString(context.Attr<std::string>("subgraph"));
-  max_batch_ = context.Attr<int>("max_batch");
+  int max_batch = context.Attr<int>("max_batch");
   auto max_workspace = context.Attr<int>("max_workspace");
-  engine_ = Singleton<TRT_EngineManager>::Global().Create(
-      max_batch_, max_workspace, &stream_);
-  engine_->InitNetwork();
+  auto params = context.Attr<std::vector<std::string>>("parameters");
+  std::unordered_set<std::string> parameters;
+  for (const auto &param : params) {
+    parameters.insert(param);
+  }
+
+  // TODO(Superjomn) replace this with a different stream
+  auto *engine = Singleton<TRT_EngineManager>::Global().Create(
+      max_batch, max_workspace, nullptr /*engine hold its own stream*/,
+      context.Attr<std::string>("engine_uniq_key"));
+  engine->InitNetwork();
 
   framework::BlockDesc block(nullptr /*programdesc*/, &block_desc);
   // Add inputs
@@ -87,24 +95,23 @@ void paddle::operators::TensorRTEngineKernel<DeviceContext, T>::Prepare(
     PADDLE_ENFORCE_EQ(var->GetType(), FluidDT::VarType_Type_LOD_TENSOR,
                       "TensorRT engine only takes LoDTensor as input");
     auto shape = var->GetShape();
-    engine_->DeclareInput(
+    engine->DeclareInput(
         input, FluidDataType2TRT(
                    var->Proto()->type().lod_tensor().tensor().data_type()),
         Vec2TRT_Dims(var->GetShape()));
   }
 
-  // TODO(Superjomn) parameters should be passed after analysised from outside.
   inference::Singleton<inference::tensorrt::OpConverter>::Global().ConvertBlock(
-      block_desc, {}, context.scope(), engine_);
+      block_desc, parameters, context.scope(), engine);
 
   // Add outputs
   VLOG(4) << "declare outputs";
   for (auto &output : context.Outputs("Ys")) {
     VLOG(4) << "declare output " << output;
-    engine_->DeclareOutput(output);
+    engine->DeclareOutput(output);
   }
 
-  engine_->FreezeNetwork();
+  engine->FreezeNetwork();
 }
 
 class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -113,6 +120,7 @@ class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("Xs", "A list of inputs.").AsDuplicable();
     AddOutput("Ys", "A list of outputs").AsDuplicable();
     AddAttr<std::string>("subgraph", "the subgraph.");
+    AddAttr<std::string>("engine_uniq_key", "unique key for the TRT engine.");
     AddAttr<int>("max_batch", "the maximum batch size.");
     AddAttr<int>("max_workspace", "the maximum batch size.");
     AddComment("TensorRT engine operator.");

paddle/fluid/operators/tensorrt_engine_op.h

@@ -19,10 +19,14 @@
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/inference/analysis/helper.h"
 #include "paddle/fluid/inference/tensorrt/engine.h"
+#include "paddle/fluid/inference/tensorrt/engine.h"
 
 namespace paddle {
 namespace operators {
 
+using inference::Singleton;
+using inference::tensorrt::TRT_EngineManager;
+
 class TensorRTEngineOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -47,16 +51,18 @@ template <typename DeviceContext, typename T>
 class TensorRTEngineKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    if (!engine_) {
+    auto engine_name = context.Attr<std::string>("engine_uniq_key");
+    if (!Singleton<TRT_EngineManager>::Global().HasEngine(engine_name)) {
       Prepare(context);
     }
+    auto* engine = Singleton<TRT_EngineManager>::Global().Get(engine_name);
     auto input_names = context.op().Inputs("Xs");
     PADDLE_ENFORCE(!input_names.empty(), "should pass more than one inputs");
     // Try to determine a batch_size
     auto& tensor0 = inference::analysis::GetFromScope<framework::LoDTensor>(
         context.scope(), input_names.front());
     int batch_size = tensor0.dims()[0];
-    PADDLE_ENFORCE_LE(batch_size, max_batch_);
+    PADDLE_ENFORCE_LE(batch_size, context.Attr<int>("max_batch"));
 
     // Convert input tensor from fluid to engine.
     for (const auto& x : context.Inputs("Xs")) {
@@ -64,20 +70,20 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
       auto& t = inference::analysis::GetFromScope<framework::LoDTensor>(
           context.scope(), x);
       if (platform::is_cpu_place(t.place())) {
-        engine_->SetInputFromCPU(x, static_cast<const void*>(t.data<void>()),
+        engine->SetInputFromCPU(x, static_cast<const void*>(t.data<void>()),
                                 t.memory_size());
       } else {
-        engine_->SetInputFromGPU(x, static_cast<const void*>(t.data<void>()),
+        engine->SetInputFromGPU(x, static_cast<const void*>(t.data<void>()),
                                 t.memory_size());
       }
     }
     // Execute the engine.
     PADDLE_ENFORCE_GT(batch_size, 0);
-    engine_->Execute(batch_size);
+    engine->Execute(batch_size);
     // Convert output tensor from engine to fluid
     for (const auto& y : context.Outputs("Ys")) {
       // convert output and copy to fluid.
-      nvinfer1::ITensor* trt_t = engine_->GetITensor(y);
+      nvinfer1::ITensor* trt_t = engine->GetITensor(y);
       auto dims = trt_t->getDimensions();
       // Use the output ITensor's dims to reshape the Fluid Tensor.
       std::vector<int> ddim(dims.d, dims.d + dims.nbDims);
@@ -89,27 +95,22 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
       auto size = inference::analysis::AccuDims(dims.d, dims.nbDims);
       if (platform::is_cpu_place(fluid_t->place())) {
         // TODO(Superjomn) change this float to dtype size.
-        engine_->GetOutputInCPU(
+        engine->GetOutputInCPU(
             y, fluid_t->mutable_data<float>(platform::CPUPlace()),
             size * sizeof(float));
       } else {
-        engine_->GetOutputInGPU(
+        engine->GetOutputInGPU(
             y, fluid_t->mutable_data<float>(platform::CUDAPlace()),
             size * sizeof(float));
       }
     }
 
-    cudaStreamSynchronize(stream_);
+    cudaStreamSynchronize(*engine->stream());
   }
 
  protected:
   // Build the engine.
   void Prepare(const framework::ExecutionContext& context) const;
-
- private:
-  mutable cudaStream_t stream_;
-  mutable inference::tensorrt::TensorRTEngine* engine_{nullptr};
-  mutable int max_batch_{0};
 };
 
 }  // namespace operators

paddle/fluid/operators/tensorrt_engine_op_test.cc

@@ -79,6 +79,17 @@ void SetAttr<int64_t>(framework::proto::OpDesc* op, const std::string& name,
   attr->set_type(paddle::framework::proto::AttrType::LONG);
   attr->set_l(data);
 }
+template <>
+void SetAttr<std::vector<std::string>>(framework::proto::OpDesc* op,
+                                       const std::string& name,
+                                       const std::vector<std::string>& data) {
+  auto* attr = op->add_attrs();
+  attr->set_name(name);
+  attr->set_type(paddle::framework::proto::AttrType::STRINGS);
+  for (const auto& s : data) {
+    attr->add_strings(s.c_str());
+  }
+}
 
 }  // namespace
 
@@ -123,11 +134,15 @@ TEST(TensorRTEngineOp, manual) {
   engine_op_desc.SetOutput("Ys", std::vector<std::string>({"z0"}));
   SetAttr<std::string>(engine_op_desc.Proto(), "subgraph",
                        block_->SerializeAsString());
-  SetAttr<int>(engine_op_desc.Proto(), "max_batch", 30);
+  SetAttr<int>(engine_op_desc.Proto(), "max_batch", 100);
   SetAttr<int>(engine_op_desc.Proto(), "max_workspace", 1 << 10);
+  SetAttr<std::string>(engine_op_desc.Proto(), "engine_uniq_key", "a_engine");
+  SetAttr<std::vector<std::string>>(engine_op_desc.Proto(), "parameters",
+                                    std::vector<std::string>({}));
 
   LOG(INFO) << "create engine op";
   auto engine_op = framework::OpRegistry::CreateOp(*engine_op_desc.Proto());
+  LOG(INFO) << "engine_op " << engine_op.get();
 
   framework::Scope scope;
   platform::CPUPlace place;
@@ -145,6 +160,88 @@ TEST(TensorRTEngineOp, manual) {
   engine_op->Run(scope, place);
 }
 
+void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) {
+  framework::ProgramDesc program;
+  framework::Scope scope;
+  platform::CPUPlace place;
+  platform::CPUDeviceContext ctx(place);
+
+  auto* block_ = program.Proto()->add_blocks();
+  block_->set_idx(0);
+  block_->set_parent_idx(-1);
+
+  using shape_t = std::vector<int64_t>;
+
+  LOG(INFO) << "create block desc";
+  framework::BlockDesc block_desc(&program, block_);
+
+  auto AddFCLayer = [&](const std::string& x_name, const std::string& y_name,
+                        const std::string& z_name, bool x_created,
+                        const shape_t& x_shape, const shape_t& y_shape,
+                        const shape_t& z_shape) {
+
+    LOG(INFO) << "create fc op";
+    auto* fc = block_desc.AppendOp();
+    fc->SetType("mul");
+    fc->SetInput("X", std::vector<std::string>({x_name}));
+    fc->SetInput("Y", std::vector<std::string>({y_name}));
+    fc->SetOutput("Out", std::vector<std::string>({z_name}));
+
+    // Set inputs' variable shape in BlockDesc
+    if (!x_created) {
+      AddTensorToBlockDesc(block_, x_name,
+                           std::vector<int64_t>({batch_size, input_dim, 1, 1}));
+    }
+    AddTensorToBlockDesc(block_, y_name,
+                         std::vector<int64_t>({input_dim, output_dim}));
+    AddTensorToBlockDesc(block_, z_name,
+                         std::vector<int64_t>({batch_size, output_dim}));
+
+    // Prepare variables.
+    if (!x_created) {
+      CreateCPUTensor(&scope, x_name, std::vector<int64_t>(x_shape));
+    }
+    CreateCPUTensor(&scope, y_name, std::vector<int64_t>(y_shape));
+    CreateCPUTensor(&scope, z_name, std::vector<int64_t>(z_shape));
+
+    // It is wired, need to copy manually.
+    *block_->add_ops() = *fc->Proto();
+  };
+
+  // Test with 4 layer FC
+  AddFCLayer("x0", "y0", "z0", false, {batch_size, input_dim},
+             {input_dim, output_dim}, {batch_size, output_dim});
+  AddFCLayer("z0", "y1", "z1", true, {}, {output_dim, output_dim},
+             {batch_size, output_dim});
+  AddFCLayer("z1", "y2", "z2", true, {}, {output_dim, output_dim},
+             {batch_size, output_dim});
+  AddFCLayer("z2", "y3", "z3", true, {}, {output_dim, output_dim},
+             {batch_size, output_dim});
+
+  LOG(INFO) << "create tensorrt desc";
+  framework::OpDesc engine_op_desc(nullptr);
+  engine_op_desc.SetType("tensorrt_engine");
+  engine_op_desc.SetInput("Xs", std::vector<std::string>({"x0"}));
+  engine_op_desc.SetOutput("Ys", std::vector<std::string>({"z3"}));
+
+  SetAttr<std::string>(engine_op_desc.Proto(), "subgraph",
+                       block_->SerializeAsString());
+  SetAttr<int>(engine_op_desc.Proto(), "max_batch", batch_size);
+  SetAttr<int>(engine_op_desc.Proto(), "max_workspace", 2 << 10);
+  SetAttr<std::vector<std::string>>(
+      engine_op_desc.Proto(), "parameters",
+      std::vector<std::string>({"y0", "y1", "y2", "y3"}));
+  SetAttr<std::string>(engine_op_desc.Proto(), "engine_uniq_key", "b_engine");
+
+  auto engine_op = framework::OpRegistry::CreateOp(*engine_op_desc.Proto());
+
+  // Execute them.
+  engine_op->Run(scope, place);
+}
+
+// Test with a larger FC layer.
+TEST(TensorRTEngineOp, fc) { Execute(40, 256, 256); }
+
 }  // namespace operators
 }  // namespace paddle
 

paddle/fluid/operators/uniform_random_op.cc

@@ -86,32 +86,24 @@ class UniformRandomOp : public framework::OperatorWithKernel {
 class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
-    AddOutput("Out", "(Tensor) The output tensor of uniform random op");
+    AddOutput("Out", "The output tensor of uniform random op");
     AddComment(R"DOC(
-Uniform random operator.
-
 This operator initializes a tensor with random values sampled from a
-uniform distribution.
+uniform distribution. The random result is in set [min, max].
 
 )DOC");
-    AddAttr<std::vector<int>>("shape",
-                              "(vector<int>) The shape of the output tensor");
-    AddAttr<float>("min",
-                   "(float, default -1.0) "
-                   "Minimum value of uniform random")
+    AddAttr<std::vector<int>>("shape", "The shape of the output tensor");
+    AddAttr<float>("min", "Minimum value of uniform random. [default -1.0].")
         .SetDefault(-1.0f);
-    AddAttr<float>("max",
-                   "(float, default 1.0) "
-                   "Maximun value of uniform random")
+    AddAttr<float>("max", "Maximun value of uniform random. [default 1.0].")
         .SetDefault(1.0f);
     AddAttr<int>("seed",
-                 "(int, default 0) "
                  "Random seed used for generating samples. "
                  "0 means use a seed generated by the system."
                  "Note that if seed is not 0, this operator will always "
-                 "generate the same random numbers every time.")
+                 "generate the same random numbers every time. [default 0].")
         .SetDefault(0);
-    AddAttr<int>("dtype", "(int, default 5(FP32)) Output tensor data type")
+    AddAttr<int>("dtype", "Output tensor data type. [default 5(FP32)].")
         .SetDefault(framework::proto::VarType::FP32);
   }
 };

python/paddle/fluid/layers/control_flow.py

@@ -20,6 +20,7 @@ from ..framework import Program, Variable, Operator
 from ..layer_helper import LayerHelper, unique_name
 from ..initializer import force_init_on_cpu
 from ops import logical_and, logical_not, logical_or
+import numpy
 
 __all__ = [
     'split_lod_tensor',
@@ -909,37 +910,40 @@ def create_array(dtype):
         dtype=dtype)
 
 
-def less_than(x, y, force_cpu=True, cond=None, **ignored):
+@templatedoc()
+def less_than(x, y, force_cpu=None, cond=None, **ignored):
     """
-    **Less than**
+    ${comment}
 
-    This layer returns the truth value of :math:`x < y` elementwise.
+    >>> import paddle.fluid as fluid
+    >>> less = fluid.layers.less_than(x=label, y=limit)
 
     Args:
-        x(Variable): First operand of *less_than*
-        y(Variable): Second operand of *less_than*
-        force_cpu(Bool|True): The output data will be on CPU if set true.
+        x(${x_type}): ${x_comment}.
+        y(${y_type}): ${y_comment}.
+        force_cpu(${force_cpu_type}): ${force_cpu_comment}.
         cond(Variable|None): Optional output variable to store the result of *less_than*
 
     Returns:
-        Variable: The tensor variable storing the output of *less_than*.
-
-    Examples:
-        .. code-block:: python
-
-          less = fluid.layers.less_than(x=label, y=limit)
+        ${out_comment}.
     """
     helper = LayerHelper("less_than", **locals())
     if cond is None:
         cond = helper.create_tmp_variable(dtype='bool')
         cond.stop_gradient = True
 
+    attrs = dict()
+    if force_cpu is not None:
+        attrs['force_cpu'] = force_cpu
+    elif force_init_on_cpu():
+        attrs['force_cpu'] = force_init_on_cpu()
+
     helper.append_op(
         type='less_than',
         inputs={'X': [x],
                 'Y': [y]},
         outputs={'Out': [cond]},
-        attrs={'force_cpu': force_cpu or force_init_on_cpu()})
+        attrs=attrs)
     return cond
 
 
@@ -1004,8 +1008,28 @@ def array_read(array, i):
 
 def shrink_memory(x, i, table):
     """
-    This function creates an operator to shrink_rnn_memory using the RankTable
+    This function creates an operator to shrink rnn memory using the RankTable
     as mentioned in the input parameter.
+
+    NOTE: This API is very low-level API. It is used by DynamicRNN only.
+
+    Since the Dynamic RNN uses no-padding way to implement RNN. The sequence
+    will be sorted by order, and the length of valid memory will be shrink after
+    each time step.
+
+    Args:
+        x(Variable): The memory object in the previous time step.
+        i(Variable): The step count variable. A int scalar as LoDTensor.
+        table(Variable): The RNNRankTable object.
+
+    Returns:
+        the memory variable after shrink.
+
+    Examples:
+
+        Since this API is very low level API. The example is not provided.
+        Please reference the implementation of class DynamicRNN for detail
+        usage.
     """
     helper = LayerHelper('shrink_memory', **locals())
     out = helper.create_tmp_variable(dtype=x.dtype)
@@ -1209,6 +1233,34 @@ class IfElseBlockGuard(object):
 
 
 class IfElse(object):
+    """
+    if-else control flow.
+
+    Args:
+        cond (Variable): condition used to compare.
+        name (str, default None): The name of this layer.
+
+    Examples:
+          .. code-block:: python
+
+            limit = fluid.layers.fill_constant_batch_size_like(
+                input=label, dtype='int64', shape=[1], value=5.0)
+            cond = fluid.layers.less_than(x=label, y=limit)
+            ie = fluid.layers.IfElse(cond)
+            with ie.true_block():
+                true_image = ie.input(image)
+                hidden = fluid.layers.fc(input=true_image, size=100, act='tanh')
+                prob = fluid.layers.fc(input=hidden, size=10, act='softmax')
+                ie.output(prob)
+
+            with ie.false_block():
+                false_image = ie.input(image)
+                hidden = fluid.layers.fc(
+                    input=false_image, size=200, act='tanh')
+                prob = fluid.layers.fc(input=hidden, size=10, act='softmax')
+                ie.output(prob)
+            prob = ie()
+    """
     OUT_IF_ELSE_BLOCKS = 0
     IN_IF_ELSE_TRUE_BLOCKS = 1
     IN_IF_ELSE_FALSE_BLOCKS = 2
@@ -1311,6 +1363,38 @@ class IfElse(object):
 
 
 class DynamicRNN(object):
+    """
+    The dynamic RNN can process a batch of sequence data. The length of each
+    sample sequence can be different. This API automatically process them in
+    batch.
+
+    The input lod must be set. Please reference `lod_tensor`
+
+    >>> import paddle.fluid as fluid
+    >>> data = fluid.layers.data(name='sentence', dtype='int64', lod_level=1)
+    >>> embedding = fluid.layers.embedding(input=data, size=[65535, 32],
+    >>>                                    is_sparse=True)
+    >>>
+    >>> drnn = fluid.layers.DynamicRNN()
+    >>> with drnn.block():
+    >>>     word = drnn.step_input(embedding)
+    >>>     prev = drnn.memory(shape=[200])
+    >>>     hidden = fluid.layers.fc(input=[word, prev], size=200, act='relu')
+    >>>     drnn.update_memory(prev, hidden)  # set prev to hidden
+    >>>     drnn.output(hidden)
+    >>>
+    >>> # last is the last time step of rnn. It is the encoding result.
+    >>> last = fluid.layers.sequence_last_step(drnn())
+
+    The dynamic RNN will unfold sequence into timesteps. Users need to define
+    how to process each time step during the :code:`with` block.
+
+    The `memory` is used staging data cross time step. The initial value of
+    memory can be zero or another variable.
+
+    The dynamic RNN can mark multiple variables as its output. Use `drnn()` to
+    get the output sequence.
+    """
     BEFORE_RNN = 0
     IN_RNN = 1
     AFTER_RNN = 2
@@ -1333,6 +1417,15 @@ class DynamicRNN(object):
         self.mem_link = []
 
     def step_input(self, x):
+        """
+        Mark a sequence as a dynamic RNN input.
+        Args:
+            x(Variable): The input sequence.
+
+        Returns:
+            The current timestep in the input sequence.
+
+        """
         self._assert_in_rnn_block_("step_input")
         if not isinstance(x, Variable):
             raise TypeError(
@@ -1376,6 +1469,15 @@ class DynamicRNN(object):
         return array_read(array=input_array, i=self.step_idx)
 
     def static_input(self, x):
+        """
+        Mark a variable as a RNN input. The input will not be scattered into
+        time steps.
+        Args:
+            x(Variable): The input variable.
+
+        Returns:
+            The input variable that can access in RNN.
+        """
         self._assert_in_rnn_block_("static_input")
         if not isinstance(x, Variable):
             raise TypeError(
@@ -1397,6 +1499,10 @@ class DynamicRNN(object):
 
     @contextlib.contextmanager
     def block(self):
+        """
+        The block for user to define operators in RNN. See the class docstring
+        for more details.
+        """
         if self.status != DynamicRNN.BEFORE_RNN:
             raise ValueError("rnn.block() can only be invoke once")
         self.step_idx = fill_constant(
@@ -1423,6 +1529,9 @@ class DynamicRNN(object):
                     x=each_array, table=self.lod_rank_table))
 
     def __call__(self, *args, **kwargs):
+        """
+        Get the output of RNN. This API should only be invoked after RNN.block()
+        """
         if self.status != DynamicRNN.AFTER_RNN:
             raise ValueError(("Output of the dynamic RNN can only be visited "
                               "outside the rnn block."))
@@ -1437,6 +1546,70 @@ class DynamicRNN(object):
                value=0.0,
                need_reorder=False,
                dtype='float32'):
+        """
+        Create a memory variable for dynamic rnn.
+
+        If the :code:`init` is not None, :code:`memory` will be initialized by
+        this variable. The :code:`need_reorder` is used to reorder the memory as
+        the input variable. It should be set to true when the initialized memory
+        depends on the input sample.
+
+        For example,
+
+        >>> import paddle.fluid as fluid
+        >>> sentence = fluid.layers.data(
+        >>>                 name='sentence', dtype='float32', shape=[32])
+        >>> boot_memory = fluid.layers.data(
+        >>>                 name='boot', dtype='float32', shape=[10])
+        >>>
+        >>> drnn = fluid.layers.DynamicRNN()
+        >>> with drnn.block():
+        >>>     word = drnn.step_input(sentence)
+        >>>     memory = drnn.memory(init=boot_memory, need_reorder=True)
+        >>>     hidden = fluid.layers.fc(
+        >>>                 input=[word, memory], size=10, act='tanh')
+        >>>     drnn.update_memory(ex_mem=memory, new_mem=hidden)
+        >>>     drnn.output(hidden)
+        >>> rnn_output = drnn()
+
+
+        Otherwise, if :code:`shape`, :code:`value`, :code:`dtype` are set, the
+        :code:`memory` will be initialized by this :code:`value`.
+
+        For example,
+
+        >>> import paddle.fluid as fluid
+        >>> sentence = fluid.layers.data(
+        >>>                 name='sentence', dtype='float32', shape=[32])
+        >>>
+        >>> drnn = fluid.layers.DynamicRNN()
+        >>> with drnn.block():
+        >>>     word = drnn.step_input(sentence)
+        >>>     memory = drnn.memory(shape=[10], dtype='float32', value=0)
+        >>>     hidden = fluid.layers.fc(
+        >>>             input=[word, memory], size=10, act='tanh')
+        >>>     drnn.update_memory(ex_mem=memory, new_mem=hidden)
+        >>>     drnn.output(hidden)
+        >>> rnn_output = drnn()
+
+
+        Args:
+            init(Variable|None): The initialized variable.
+
+            shape(list|tuple): The memory shape. NOTE the shape does not contain
+            batch_size.
+
+            value(float): the initalized value.
+
+            need_reorder(bool): True if the initialized memory depends on the
+            input sample.
+
+            dtype(str|numpy.dtype): The data type of the initialized memory.
+
+        Returns:
+            the memory variable.
+
+        """
         self._assert_in_rnn_block_('memory')
         if init is not None:
             if not isinstance(init, Variable):
@@ -1504,6 +1677,16 @@ class DynamicRNN(object):
             return self.memory(init=init)
 
     def update_memory(self, ex_mem, new_mem):
+        """
+        Update the memory from ex_mem to new_mem. NOTE that the shape and data
+        type of :code:`ex_mem` and :code:`new_mem` must be same.
+        Args:
+            ex_mem(Variable): the memory variable.
+            new_mem(Variable): the plain variable generated in RNN block.
+
+        Returns:
+            None
+        """
         self._assert_in_rnn_block_('update_memory')
         if not isinstance(ex_mem, Variable):
             raise TypeError("The input arg `ex_mem` of update_memory() must "
@@ -1521,6 +1704,15 @@ class DynamicRNN(object):
         self.mem_link.append((new_mem, mem_array))
 
     def output(self, *outputs):
+        """
+        mark the RNN output variables.
+
+        Args:
+            outputs: The output variables.
+
+        Returns:
+            None
+        """
         self._assert_in_rnn_block_('output')
         parent_block = self._parent_block_()
         for each in outputs:

python/paddle/fluid/layers/detection.py

@@ -210,53 +210,68 @@ def bipartite_match(dist_matrix,
                     dist_threshold=None,
                     name=None):
     """
-    **Bipartite matchint operator**
-
-    This operator is a greedy bipartite matching algorithm, which is used to
-    obtain the matching with the maximum distance based on the input
+    This operator implements a greedy bipartite matching algorithm, which is
+    used to obtain the matching with the maximum distance based on the input
     distance matrix. For input 2D matrix, the bipartite matching algorithm can
-    find the matched column for each row, also can find the matched row for
-    each column. And this operator only calculate matched indices from column
-    to row. For each instance, the number of matched indices is the number of
-    of columns of the input ditance matrix.
-
-    There are two outputs to save matched indices and distance.
-    A simple description, this algothrim matched the best (maximum distance)
+    find the matched column for each row (matched means the largest distance),
+    also can find the matched row for each column. And this operator only
+    calculate matched indices from column to row. For each instance,
+    the number of matched indices is the column number of the input distance
+    matrix.
+
+    There are two outputs, matched indices and distance.
+    A simple description, this algorithm matched the best (maximum distance)
     row entity to the column entity and the matched indices are not duplicated
     in each row of ColToRowMatchIndices. If the column entity is not matched
     any row entity, set -1 in ColToRowMatchIndices.
 
-    Please note that the input DistMat can be LoDTensor (with LoD) or Tensor.
+    NOTE: the input DistMat can be LoDTensor (with LoD) or Tensor.
     If LoDTensor with LoD, the height of ColToRowMatchIndices is batch size.
     If Tensor, the height of ColToRowMatchIndices is 1.
 
+    NOTE: This API is a very low level API. It is used by :code:`ssd_loss`
+    layer. Please consider to use :code:`ssd_loss` instead.
+
     Args:
         dist_matrix(Variable): This input is a 2-D LoDTensor with shape
             [K, M]. It is pair-wise distance matrix between the entities
             represented by each row and each column. For example, assumed one
             entity is A with shape [K], another entity is B with shape [M]. The
-            dist_matirx[i][j] is the distance between A[i] and B[j]. The bigger
-            the distance is, the better macthing the pairs are. Please note,
-            This tensor can contain LoD information to represent a batch of
-            inputs. One instance of this batch can contain different numbers of
-            entities.
+            dist_matrix[i][j] is the distance between A[i] and B[j]. The bigger
+            the distance is, the better matching the pairs are.
+
+            NOTE: This tensor can contain LoD information to represent a batch
+            of inputs. One instance of this batch can contain different numbers
+            of entities.
         match_type(string|None): The type of matching method, should be
-           'bipartite' or 'per_prediction', 'bipartite' by defalut.
+           'bipartite' or 'per_prediction'. [default 'bipartite'].
         dist_threshold(float|None): If `match_type` is 'per_prediction',
             this threshold is to determine the extra matching bboxes based
-            on the maximum distance, 0.5 by defalut.
+            on the maximum distance, 0.5 by default.
     Returns:
-        match_indices(Variable): A 2-D Tensor with shape [N, M] in int type.
+        tuple: a tuple with two elements is returned. The first is
+        matched_indices, the second is matched_distance.
+
+        The matched_indices is a 2-D Tensor with shape [N, M] in int type.
         N is the batch size. If match_indices[i][j] is -1, it
         means B[j] does not match any entity in i-th instance.
         Otherwise, it means B[j] is matched to row
         match_indices[i][j] in i-th instance. The row number of
         i-th instance is saved in match_indices[i][j].
-        match_distance(Variable): A 2-D Tensor with shape [N, M] in float type.
-            N is batch size. If match_indices[i][j] is -1,
+
+        The matched_distance is a 2-D Tensor with shape [N, M] in float type
+        . N is batch size. If match_indices[i][j] is -1,
         match_distance[i][j] is also -1.0. Otherwise, assumed
         match_distance[i][j] = d, and the row offsets of each instance
-            are called LoD. Then match_distance[i][j] = dist_matrix[d+LoD[i]][j].
+        are called LoD. Then match_distance[i][j] =
+        dist_matrix[d+LoD[i]][j].
+
+    Examples:
+
+        >>> x = fluid.layers.data(name='x', shape=[4], dtype='float32')
+        >>> y = fluid.layers.data(name='y', shape=[4], dtype='float32')
+        >>> iou = fluid.layers.iou_similarity(x=x, y=y)
+        >>> matched_indices, matched_dist = fluid.layers.bipartite_match(iou)
     """
     helper = LayerHelper('bipartite_match', **locals())
     match_indices = helper.create_tmp_variable(dtype='int32')
@@ -364,7 +379,7 @@ def ssd_loss(location,
              normalize=True,
              sample_size=None):
     """
-    **Multi-box loss layer for object dection algorithm of SSD**
+    **Multi-box loss layer for object detection algorithm of SSD**
 
     This layer is to compute dection loss for SSD given the location offset
     predictions, confidence predictions, prior boxes and ground-truth boudding
@@ -372,21 +387,35 @@ def ssd_loss(location,
     is a weighted sum of the localization loss (or regression loss) and
     confidence loss (or classification loss) by performing the following steps:
 
-    1. Find matched boundding box by bipartite matching algorithm.
+    1. Find matched bounding box by bipartite matching algorithm.
+
       1.1 Compute IOU similarity between ground-truth boxes and prior boxes.
+
       1.2 Compute matched boundding box by bipartite matching algorithm.
+
     2. Compute confidence for mining hard examples
+
       2.1. Get the target label based on matched indices.
+
       2.2. Compute confidence loss.
+
     3. Apply hard example mining to get the negative example indices and update
        the matched indices.
+
     4. Assign classification and regression targets
+
       4.1. Encoded bbox according to the prior boxes.
+
       4.2. Assign regression targets.
+
       4.3. Assign classification targets.
+
     5. Compute the overall objective loss.
+
       5.1 Compute confidence loss.
+
       5.1 Compute localization loss.
+
       5.3 Compute the overall weighted loss.
 
     Args:
@@ -421,39 +450,36 @@ def ssd_loss(location,
         mining_type (str): The hard example mining type, should be 'hard_example'
             or 'max_negative', now only support `max_negative`.
         normalize (bool): Whether to normalize the SSD loss by the total number
-            of output locations, True by defalut.
+            of output locations, True by default.
         sample_size (int): The max sample size of negative box, used only when
             mining_type is 'hard_example'.
 
     Returns:
-        Variable: The weighted sum of the localization loss and confidence loss,
-            with shape [N * Np, 1], N and Np are the same as they are
-            in `location`.
+        The weighted sum of the localization loss and confidence loss, with \
+        shape [N * Np, 1], N and Np are the same as they are in `location`.
 
     Raises:
-        ValueError: If mining_type is 'hard_example', now only support
-            mining type of `max_negative`.
+        ValueError: If mining_type is 'hard_example', now only support mining \
+        type of `max_negative`.
 
     Examples:
-        .. code-block:: python
-
-            pb = layers.data(
-                name='prior_box',
-                shape=[10, 4],
-                append_batch_size=False,
-                dtype='float32')
-            pbv = layers.data(
-                name='prior_box_var',
-                shape=[10, 4],
-                append_batch_size=False,
-                dtype='float32')
-            loc = layers.data(name='target_box', shape=[10, 4], dtype='float32')
-            scores = layers.data(name='scores', shape=[10, 21], dtype='float32')
-            gt_box = layers.data(
-                name='gt_box', shape=[4], lod_level=1, dtype='float32')
-            gt_label = layers.data(
-                name='gt_label', shape=[1], lod_level=1, dtype='float32')
-            loss = layers.ssd_loss(loc, scores, gt_box, gt_label, pb, pbv)
+        >>> pb = fluid.layers.data(
+        >>>                   name='prior_box',
+        >>>                   shape=[10, 4],
+        >>>                   append_batch_size=False,
+        >>>                   dtype='float32')
+        >>> pbv = fluid.layers.data(
+        >>>                   name='prior_box_var',
+        >>>                   shape=[10, 4],
+        >>>                   append_batch_size=False,
+        >>>                   dtype='float32')
+        >>> loc = fluid.layers.data(name='target_box', shape=[10, 4], dtype='float32')
+        >>> scores = fluid.layers.data(name='scores', shape=[10, 21], dtype='float32')
+        >>> gt_box = fluid.layers.data(
+        >>>         name='gt_box', shape=[4], lod_level=1, dtype='float32')
+        >>> gt_label = fluid.layers.data(
+        >>>         name='gt_label', shape=[1], lod_level=1, dtype='float32')
+        >>> loss = fluid.layers.ssd_loss(loc, scores, gt_box, gt_label, pb, pbv)
     """
 
     helper = LayerHelper('ssd_loss', **locals())

python/paddle/fluid/layers/io.py

@@ -292,6 +292,7 @@ def _copy_reader_create_op_(block, op):
     return new_op
 
 
+@templatedoc(op_type='create_recordio_file_reader')
 def open_recordio_file(filename,
                        shapes,
                        lod_levels,
@@ -299,34 +300,30 @@ def open_recordio_file(filename,
                        pass_num=1,
                        for_parallel=True):
     """
-    Open a RecordIO file
-
-    This layer takes a RecordIO file to read from and returns a Reader Variable.
-    Via the Reader Variable, we can get data from the given RecordIO file.
+    ${comment}
 
     Args:
-       filename(str): The RecordIO file's name.
+       filename(${filename_type}): ${filename_comment}.
        shapes(list): List of tuples which declaring data shapes.
-       lod_levels(list): List of ints which declaring data lod_level.
+       lod_levels(${lod_levels_type}): ${lod_levels_comment}.
        dtypes(list): List of strs which declaring data type.
        pass_num(int): Number of passes to run.
        for_parallel(Bool): Set it as True if you are going to run
             subsequent operators in parallel.
 
     Returns:
-       Variable: A Reader Variable via which we can get RecordIO file data.
+       ${out_comment}.
 
     Examples:
-       .. code-block:: python
 
-         reader = fluid.layers.io.open_recordio_file(
-                                          filename='./data.recordio',
-                                          shapes=[(3,224,224), (1)],
-                                          lod_levels=[0, 0],
-                                          dtypes=['float32', 'int64'])
-
-         # Via the reader, we can use 'read_file' layer to get data:
-         image, label = fluid.layers.io.read_file(reader)
+        >>> import paddle.fluid as fluid
+        >>> reader = fluid.layers.io.open_recordio_file(
+        >>>                               filename='./data.recordio',
+        >>>                               shapes=[(3,224,224), (1)],
+        >>>                               lod_levels=[0, 0],
+        >>>                               dtypes=['float32', 'int64'])
+        >>> # Via the reader, we can use 'read_file' layer to get data:
+        >>> image, label = fluid.layers.io.read_file(reader)
     """
     dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes]
     shape_concat = []
@@ -544,6 +541,9 @@ def __create_unshared_decorated_reader__(op_type, reader, attrs, name=None):
 
 
 def shuffle(reader, buffer_size):
+    """
+    Shuffle the reader.
+    """
     return __create_unshared_decorated_reader__(
         'create_shuffle_reader', reader, {'buffer_size': int(buffer_size)})
 
@@ -554,6 +554,29 @@ def batch(reader, batch_size):
 
 
 def double_buffer(reader, place=None, name=None):
+    """
+    Wrap a double buffer reader. The data will copy to target place with a
+    double buffer queue. If the target place is None, the place that executor
+    perform on will be used.
+
+    Args:
+        reader(Variable): the reader variable need to be wrapped.
+        place(Place): the place of target data. Default is the sample place of
+            executor perform.
+
+        name(str): Variable name. None if the user does not care.
+
+    Returns:
+        wrapped reader with double buffer.
+
+    Examples:
+
+        >>> reader = fluid.layers.open_files(filenames=['somefile'],
+        >>>                                  shapes=[[-1, 784], [-1, 1]],
+        >>>                                  dtypes=['float32', 'int64'])
+        >>> reader = fluid.layers.double_buffer(reader)
+        >>> img, label = fluid.layers.read_file(reader)
+    """
     attrs = dict()
     if place is not None:
         attrs['place'] = str(place).upper()
@@ -587,6 +610,26 @@ def read_file(file_obj):
 
 
 class Preprocessor(object):
+    """
+    A block for data pre-processing in reader.
+
+    Args:
+        reader (Variable): A reader variable.
+        name (str, default None): The name of the reader.
+
+    Examples:
+          .. code-block:: python
+
+            preprocessor = fluid.layers.io.Preprocessor(reader=reader)
+            with preprocessor.block():
+                img, lbl = preprocessor.inputs()
+                img_out = img / 2
+                lbl_out = lbl + 1
+                preprocessor.outputs(img_out, lbl_out)
+
+            data_file = fluid.layers.io.double_buffer(preprocessor())
+
+    """
     BEFORE_SUB_BLOCK = 0
     IN_SUB_BLOCK = 1
     AFTER_SUB_BLOCK = 2

python/paddle/fluid/layers/ops.py

@@ -38,8 +38,6 @@ __activations__ = [
     'relu6',
     'pow',
     'stanh',
-    'hard_shrink',
-    'thresholded_relu',
     'hard_sigmoid',
     'swish',
 ]
@@ -62,11 +60,9 @@ __all__ = [
     'logical_or',
     'logical_xor',
     'logical_not',
-    'uniform_random',
     'uniform_random_batch_size_like',
     'gaussian_random',
     'gaussian_random_batch_size_like',
-    'cumsum',
     'scatter',
     'sum',
     'slice',
@@ -77,3 +73,88 @@ __all__ = [
 
 for _OP in set(__all__):
     globals()[_OP] = generate_layer_fn(_OP)
+
+__all__ += ["uniform_random"]
+
+_uniform_random_ = generate_layer_fn('uniform_random')
+
+
+def uniform_random(shape, dtype=None, min=None, max=None, seed=None):
+    kwargs = dict()
+    for name in locals():
+        val = locals()[name]
+        if val is not None:
+            kwargs[name] = val
+    return _uniform_random_(**kwargs)
+
+
+uniform_random.__doc__ = _uniform_random_.__doc__ + """
+Examples:
+
+    >>> result = fluid.layers.uniform_random(shape=[32, 784])
+"""
+
+__all__ += ['hard_shrink']
+
+_hard_shrink_ = generate_layer_fn('hard_shrink')
+
+
+def hard_shrink(x, threshold=None):
+    kwargs = dict()
+    for name in locals():
+        val = locals()[name]
+        if val is not None:
+            kwargs[name] = val
+    return _hard_shrink_(**kwargs)
+
+
+hard_shrink.__doc__ = _hard_shrink_.__doc__ + """
+Examples:
+
+    >>> data = fluid.layers.data(name="input", shape=[784])
+    >>> result = fluid.layers.hard_shrink(x=data, threshold=0.3)
+"""
+
+__all__ += ['cumsum']
+
+_cum_sum_ = generate_layer_fn('cumsum')
+
+
+def cumsum(x, axis=None, exclusive=None, reverse=None):
+    kwargs = dict()
+    for name in locals():
+        val = locals()[name]
+        if val is not None:
+            kwargs[name] = val
+
+    return _cum_sum_(**kwargs)
+
+
+cumsum.__doc__ = _cum_sum_.__doc__ + """
+Examples:
+
+    >>> data = fluid.layers.data(name="input", shape=[32, 784])
+    >>> result = fluid.layers.cumsum(data, axis=0)
+"""
+
+__all__ += ['thresholded_relu']
+
+_thresholded_relu_ = generate_layer_fn('thresholded_relu')
+
+
+def thresholded_relu(x, threshold=None):
+    kwargs = dict()
+    for name in locals():
+        val = locals()[name]
+        if val is not None:
+            kwargs[name] = val
+
+    _thresholded_relu_(**kwargs)
+
+
+thresholded_relu.__doc__ = _thresholded_relu_.__doc__ + """
+Examples:
+
+    >>> data = fluid.layers.data(name="input", shape=[1])
+    >>> result = fluid.layers.thresholded_relu(data, threshold=0.4)
+"""

python/paddle/fluid/layers/tensor.py

@@ -6,7 +6,7 @@
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
-# Unless required by applicable law or agreed to in writing, software
+# Unlessf 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
@@ -51,7 +51,12 @@ def create_parameter(shape,
                      is_bias=False,
                      default_initializer=None):
     """
-    Create a parameter
+    Create a parameter. The parameter is a learnable variable, which can have
+    gradient, and can be optimized.
+
+    NOTE: this is a very low-level API. This API is useful when you create
+    operator by your self. instead of using layers.
+
     Args:
         shape(list[int]): shape of the parameter
         dtype(string): element type of the parameter
@@ -63,7 +68,12 @@ def create_parameter(shape,
         default_initializer(Initializer): initializer for the parameter
 
     Returns:
-        Parameter: the created parameter
+        the created parameter.
+
+    Examples:
+        >>> W = fluid.layers.create_parameter(shape=[784, 200], dtype='float32')
+        >>> data = fluid.layers.data(name="img", shape=[64, 784], append_batch_size=False)
+        >>> hidden = fluid.layers.matmul(x=data, y=W)
     """
     helper = LayerHelper("create_parameter", **locals())
     if attr is None:
@@ -191,6 +201,7 @@ def assign(input, output):
 
     Examples:
         .. code-block:: python
+
           out = fluid.layers.create_tensor(dtype='float32')
           hidden = fluid.layers.fc(input=data, size=10)
           fluid.layers.assign(hidden, out)

python/paddle/fluid/tests/book/test_label_semantic_roles.py

@@ -76,8 +76,7 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark,
     emb_layers.append(mark_embedding)
 
     hidden_0_layers = [
-        fluid.layers.fc(input=emb, size=hidden_dim, act='tanh')
-        for emb in emb_layers
+        fluid.layers.fc(input=emb, size=hidden_dim) for emb in emb_layers
     ]
 
     hidden_0 = fluid.layers.sums(input=hidden_0_layers)
@@ -94,8 +93,8 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark,
 
     for i in range(1, depth):
         mix_hidden = fluid.layers.sums(input=[
-            fluid.layers.fc(input=input_tmp[0], size=hidden_dim, act='tanh'),
-            fluid.layers.fc(input=input_tmp[1], size=hidden_dim, act='tanh')
+            fluid.layers.fc(input=input_tmp[0], size=hidden_dim),
+            fluid.layers.fc(input=input_tmp[1], size=hidden_dim)
         ])
 
         lstm = fluid.layers.dynamic_lstm(

python/paddle/fluid/transpiler/memory_optimization_transpiler.py

@@ -157,9 +157,11 @@ class ControlFlowGraph(object):
             if op.type() == "fill_constant" and op.attr("force_cpu") == True:
                 self._skip_opt.update(op.output_arg_names())
 
-    def release_memory(self):
+    def release_memory(self, skip_opt_set=None):
         self._dataflow_analyze()
         self._update_skip_opt_set()
+        if skip_opt_set:
+            self._skip_opt.update(skip_opt_set)
         fwd_id = 0
         bwd_id = 0
         for i in range(self.op_size):
@@ -183,7 +185,7 @@ class ControlFlowGraph(object):
                 else:
                     bwd_id += 1
 
-    def memory_optimize(self, level=0):
+    def memory_optimize(self, skip_opt_set=None, level=0):
         def compare_shape(x_shape, cache_shape, opt_level):
             if opt_level == 0:
                 return x_shape == cache_shape
@@ -200,6 +202,9 @@ class ControlFlowGraph(object):
 
         self._dataflow_analyze()
         self._update_skip_opt_set()
+        # update skip set to meet users' demand
+        if skip_opt_set:
+            self._skip_opt.update(skip_opt_set)
         self.pool = []
         for i in range(self.op_size):
             op = self._ops[i]
@@ -358,7 +363,7 @@ def _get_cfgs(input_program):
     return cfgs
 
 
-def memory_optimize(input_program, print_log=False, level=0):
+def memory_optimize(input_program, skip_opt_set=None, print_log=False, level=0):
     """Optimize memory by reusing var memory.
 
       Note: it doesn't not support subblock nested in subblock.
@@ -374,10 +379,10 @@ def memory_optimize(input_program, print_log=False, level=0):
     PRINT_LOG = print_log
     cfgs = _get_cfgs(input_program)
     for cfg in cfgs:
-        cfg.memory_optimize(level)
+        cfg.memory_optimize(skip_opt_set=skip_opt_set, level=level)
 
 
-def release_memory(input_program):
+def release_memory(input_program, skip_opt_set=None):
     cfgs = _get_cfgs(input_program)
     for cfg in cfgs:
-        cfg.release_memory()
+        cfg.release_memory(skip_opt_set=skip_opt_set)