Merge branch 'develop' of https://github.com/paddlepaddle/paddle into restructure-attribute

CMakeLists.txt

@@ -36,8 +36,8 @@ include(simd)
 ################################ Configurations #######################################
 option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_FOUND})
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
-option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    ${AVX_FOUND})
-option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      ${AVX_FOUND})
+option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    OFF)
+option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      OFF)
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        ON)
 option(WITH_SWIG_PY     "Compile PaddlePaddle with inference api"       ON)

cmake/generic.cmake

@@ -187,7 +187,13 @@ function(cc_library TARGET_NAME)
     endif()
     
     # cpplint code style
-    add_style_check_target(${TARGET_NAME} ${cc_library_SRCS})
+    foreach(source_file ${cc_library_SRCS})
+      string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file})
+      if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+        list(APPEND cc_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+      endif()
+    endforeach()
+    add_style_check_target(${TARGET_NAME} ${cc_library_SRCS} ${cc_library_HEADERS})
 
   else(cc_library_SRCS)
     if (cc_library_DEPS)
@@ -239,6 +245,14 @@ function(nv_library TARGET_NAME)
         add_dependencies(${TARGET_NAME} ${nv_library_DEPS})
         target_link_libraries(${TARGET_NAME} ${nv_library_DEPS})
       endif()
+      # cpplint code style
+      foreach(source_file ${nv_library_SRCS})
+        string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file})
+        if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+          list(APPEND cc_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+        endif()
+      endforeach()
+      add_style_check_target(${TARGET_NAME} ${nv_library_SRCS} ${nv_library_HEADERS})
     else(nv_library_SRCS)
       if (nv_library_DEPS)
         merge_static_libs(${TARGET_NAME} ${nv_library_DEPS})

paddle/framework/ddim.h

@@ -25,18 +25,15 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-namespace {
-typedef boost::variant<Dim<1>, Dim<2>, Dim<3>, Dim<4>, Dim<5>, Dim<6>, Dim<7>,
-                       Dim<8>, Dim<9>>
-    DDimVar;
-}
-
 /**
  * \brief A dynamically sized dimension.
  *
  * The number of dimensions must be between [1, 9].
  */
 struct DDim {
+  typedef boost::variant<Dim<1>, Dim<2>, Dim<3>, Dim<4>, Dim<5>, Dim<6>, Dim<7>,
+                         Dim<8>, Dim<9>>
+      DDimVar;
   DDimVar var;
 
   DDim() : var(Dim<1>()) {}

paddle/framework/grad_op_builder.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
 #pragma once
 
 #include "paddle/framework/operator.h"

paddle/framework/op_registry.h

@@ -271,7 +271,7 @@ class OpRegistry {
   static std::unordered_map<std::string, OpProto>& protos() {
     static std::unordered_map<std::string, OpProto> protos_;
     return protos_;
-  };
+  }
 
   static std::unordered_map<std::string, std::string>& grad_ops() {
     static std::unordered_map<std::string, std::string> grad_ops_;
@@ -293,7 +293,7 @@ class OpRegistry {
   static std::unordered_map<std::string, OpAttrChecker>& op_checkers() {
     static std::unordered_map<std::string, OpAttrChecker> op_checkers_;
     return op_checkers_;
-  };
+  }
 
   static void GenerateTempVariableName(OperatorBase* op) {
     static std::atomic<size_t> gUniqId(0UL);
@@ -310,7 +310,7 @@ class OpRegistry {
 template <typename OpType, typename ProtoMakerType>
 class OpRegisterHelper {
  public:
-  OpRegisterHelper(const char* op_type) {
+  explicit OpRegisterHelper(const char* op_type) {
     OpRegistry::RegisterOp<OpType, ProtoMakerType>(op_type);
   }
 };

paddle/framework/operator.h

@@ -284,7 +284,7 @@ class OperatorWithKernel : public OperatorBase {
     platform::Place place_;
 
     OpKernelKey() = default;
-    OpKernelKey(const platform::DeviceContext& dev_ctx) {
+    explicit OpKernelKey(const platform::DeviceContext& dev_ctx) {
       place_ = dev_ctx.GetPlace();
     }
 

paddle/framework/pybind.cc

@@ -105,7 +105,16 @@ PYBIND11_PLUGIN(core) {
       .def("set", PyCUDATensorSetFromArray<float>)
       .def("set", PyCUDATensorSetFromArray<int>)
 #endif
-      .def("shape", [](Tensor &self) { return vectorize(self.dims()); });
+      .def("shape", [](Tensor &self) { return vectorize(self.dims()); })
+      .def("set_float_element",
+           [](Tensor &self, size_t offset, float f) {
+             // TODO(yuyang18): Only support GPU now.
+             self.data<float>()[offset] = f;
+           })
+      .def("get_float_element", [](Tensor &self, size_t offset) -> float {
+        // TODO(yuyang18): Only support GPU now.
+        return self.data<float>()[offset];
+      });
 
   py::class_<Variable>(m, "Variable", R"DOC(Variable Class.
 

paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp

@@ -967,8 +967,9 @@ void RecurrentGradientMachine::generateSequence() {
   size_t numSequences = getGenBatchSize();
 
   resizeBootFrame(numSequences);
-  // We create only two sub-network in generation for alternate use.
-  // Thus, we can reduce total memory of output_ in layer forward.
+  // We create only two sub-network in generation, one stores states of all
+  // layers in previous time step and the other storing the states at current
+  // time step.
   resizeOrCreateFrames(2);
 
   // outFrameLines_.size() > 1UL
@@ -1001,10 +1002,9 @@ void RecurrentGradientMachine::generateSequence() {
 
   // init outArg
   size_t resultNum = generator_.config.num_results_per_sample();
-  IVector::resizeOrCreate(
-      generator_.outArg.ids,
-      generator_.config.max_num_frames() * numSequences * resultNum,
-      false);
+  size_t maxGenWordCount =
+      generator_.config.max_num_frames() * numSequences * resultNum;
+  IVector::resizeOrCreate(generator_.outArg.ids, maxGenWordCount, false);
   if (resultNum > 1) {
     CHECK_LE(resultNum, static_cast<size_t>(generator_.config.beam_size()));
     Matrix::resizeOrCreate(generator_.outArg.in,
@@ -1012,6 +1012,11 @@ void RecurrentGradientMachine::generateSequence() {
                            /* width */ resultNum,
                            false,
                            /* useGpu */ false);
+    Matrix::resizeOrCreate(generator_.outArg.value,
+                           /* height */ maxGenWordCount,
+                           /* width */ 1,
+                           false,
+                           /* useGpu */ false);
   }
   ICpuGpuVector::resizeOrCreate(generator_.outArg.sequenceStartPositions,
                                 numSequences + 1,
@@ -1313,13 +1318,20 @@ void RecurrentGradientMachine::fillGenOutputs() {
   starts[0] = 0;
   if (numResults > 1) {
     real* probs = generator_.outArg.in->getData();
+    real* idsProb = generator_.outArg.value->getData();
+    size_t curPos = 0;
     for (size_t i = 0; i < finalPaths_.size(); ++i) {
       for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
         Path& path = finalPaths_[i][j];
-        generator_.ids.push_back(path.ids.size());  // sequence size
+        size_t genLen = path.ids.size();
+        generator_.ids.push_back(genLen);  // sequence size
         generator_.ids.insert(
             generator_.ids.end(), path.ids.begin(), path.ids.end());
         generator_.ids.push_back(-1);  // end of sequence
+
+        memcpy(idsProb + curPos, path.idsProb.data(), sizeof(real) * genLen);
+        curPos += genLen;
+        idsProb[curPos++] = -1.0;
         probs[i * numResults + j] = path.logProb;
 
         if (!j && dataArgsSize_) {

paddle/gserver/gradientmachines/RecurrentGradientMachine.h

@@ -189,6 +189,11 @@ public:
      */
     std::vector<int> ids;
 
+    /**
+     * @brief idsProb, log probability of each generated words.
+     */
+    std::vector<real> idsProb;
+
     /**
      * @brief logProb, current probability of path.
      */
@@ -228,11 +233,13 @@ public:
      */
     Path(Path& old, int newId, real logProb, int machineId, int topIndex)
         : ids(old.ids),
+          idsProb(old.idsProb),
           logProb(old.logProb + logProb),
           machineId(machineId),
           topIndex(topIndex),
           seqId(old.seqId) {
       ids.push_back(newId);
+      idsProb.push_back(logProb);
       if (!old.probHistory.empty()) {
         this->probHistory = old.probHistory;
         // probHistory store current prob, not sum
@@ -411,8 +418,9 @@ protected:
 
   struct Generator {
     GeneratorConfig config;
-    std::vector<int> ids;  // store generated sequences
-    Argument outArg;       // final output argument
+    std::vector<int> ids;       // store generated sequences
+    std::vector<real> idsProb;  // log probability of each generated word
+    Argument outArg;            // final output argument
   };
   bool generating_;
   Generator generator_;

paddle/gserver/tests/LayerGradUtil.cpp

@@ -400,7 +400,6 @@ void initDataLayer(TestConfig testConf,
         const std::vector<int>& labelSeqStartPositions =
             testConf.inputDefs[i].labelSeqStartPositions;
         if (labelSeqStartPositions.size() != 0) {
-          CHECK(!sequenceStartPositions);
           CHECK_GE(static_cast<int>(labelSeqStartPositions.size()), 2);
 
           sequenceStartPositions =
@@ -410,6 +409,19 @@ void initDataLayer(TestConfig testConf,
                                            useGpu);
           data.sequenceStartPositions = sequenceStartPositions;
         }
+
+        const std::vector<int>& labelSubSeqStartPositions =
+            testConf.inputDefs[i].labelSubSeqStartPositions;
+        if (labelSubSeqStartPositions.size() != 0) {
+          CHECK_GE(static_cast<int>(labelSubSeqStartPositions.size()), 2);
+
+          subSequenceStartPositions =
+              ICpuGpuVector::create(labelSubSeqStartPositions.size(), useGpu);
+          subSequenceStartPositions->copyFrom(labelSubSeqStartPositions.data(),
+                                              labelSubSeqStartPositions.size(),
+                                              useGpu);
+          data.subSequenceStartPositions = subSequenceStartPositions;
+        }
         break;
       }
       default:

paddle/gserver/tests/LayerGradUtil.h

@@ -67,6 +67,7 @@ struct InputDef {
   bool isStatic;
   std::vector<int> labelInitValue;
   std::vector<int> labelSeqStartPositions;
+  std::vector<int> labelSubSeqStartPositions;
   MatrixPtr selfDefinedData;
 
   InputDef(InputType type, string nameIn, size_t dimIn, size_t sizeIn) {
@@ -81,8 +82,10 @@ struct InputDef {
   InputDef(InputType type,
            string nameIn,
            MatrixPtr selfDefinedData,
-           std::vector<int> selfDefinedSeqStartPos = {})
+           std::vector<int> selfDefinedSeqStartPos = {},
+           std::vector<int> selfDefinedSubSeqStartPos = {})
       : labelSeqStartPositions(selfDefinedSeqStartPos),
+        labelSubSeqStartPositions(selfDefinedSubSeqStartPos),
         selfDefinedData(selfDefinedData) {
     inputType = type;
     name = nameIn;

paddle/memory/detail/buddy_allocator.h

@@ -39,7 +39,7 @@ class BuddyAllocator {
 
  public:
   void* Alloc(size_t unaligned_size);
-  void Free(void*);
+  void Free(void* ptr);
   size_t Used();
 
  public:

paddle/memory/detail/meta_cache.h

@@ -33,17 +33,17 @@ namespace detail {
  */
 class MetadataCache {
  public:
-  MetadataCache(bool uses_gpu);
+  explicit MetadataCache(bool uses_gpu);
 
  public:
   /*! \brief Load the associated metadata for the specified memory block. */
-  Metadata load(const MemoryBlock*);
+  Metadata load(const MemoryBlock* memory_block);
 
   /*! \brief Store the associated metadata for the specified memory block. */
-  void store(MemoryBlock*, const Metadata&);
+  void store(MemoryBlock* memory_block, const Metadata& meta_data);
 
   /*! \brief Indicate that the specified metadata will no longer be used. */
-  void invalidate(MemoryBlock*);
+  void invalidate(MemoryBlock* memory_block);
 
  public:
   MetadataCache(const MetadataCache&) = delete;

paddle/memory/memory.h

@@ -68,7 +68,7 @@ class PODDeleter {
   static_assert(std::is_pod<T>::value, "T must be POD");
 
  public:
-  PODDeleter(Place place) : place_(place) {}
+  explicit PODDeleter(Place place) : place_(place) {}
   void operator()(T* ptr) { Free(place_, static_cast<void*>(ptr)); }
 
  private:

paddle/operators/add_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
 #define EIGEN_USE_GPU
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/add_op.h"

paddle/operators/cross_entropy_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
 #define EIGEN_USE_GPU
 #include "paddle/operators/cross_entropy_op.h"
 
 REGISTER_OP_GPU_KERNEL(onehot_cross_entropy,
-                       ops::OnehotCrossEntropyOpKernel<ops::GPUPlace, float>);
\ No newline at end of file
+                       ops::OnehotCrossEntropyOpKernel<ops::GPUPlace, float>);

paddle/operators/fill_zeros_like_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/fill_zeros_like_op.h"
 
 REGISTER_OP_GPU_KERNEL(
     fill_zeros_like,
-    paddle::operators::FillZerosLikeKernel<paddle::platform::GPUPlace, float>);
\ No newline at end of file
+    paddle::operators::FillZerosLikeKernel<paddle::platform::GPUPlace, float>);

paddle/operators/mean_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
 #define EIGEN_USE_GPU
 
 #include "paddle/operators/mean_op.h"
 
 REGISTER_OP_GPU_KERNEL(mean, ops::MeanKernel<ops::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(mean_grad, ops::MeanGradKernel<ops::GPUPlace, float>);
\ No newline at end of file
+REGISTER_OP_GPU_KERNEL(mean_grad, ops::MeanGradKernel<ops::GPUPlace, float>);

paddle/operators/mul_op.cu

@@ -15,4 +15,4 @@
 #define EIGEN_USE_GPU
 #include "paddle/operators/mul_op.h"
 
-REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<ops::GPUPlace, float>);
\ No newline at end of file
+REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<ops::GPUPlace, float>);

paddle/operators/recurrent_op.h

@@ -19,7 +19,7 @@
 namespace paddle {
 namespace operators {
 
-using namespace paddle::framework;
+using namespace paddle::framework;  // NOLINT
 
 namespace rnn {
 
@@ -94,7 +94,7 @@ void InitArgument(const ArgumentName& name, Argument* arg);
 };  // namespace rnn
 
 // The sequence format in RecurrentOp is Tensor<seq_len, batch_size, dim> now.
-// TODO:
+// TODO(Yan Chunwei):
 // 1. No-padding computing for sequences with indifinite length in one batch.
 // 2. Hierarchical RNN for sequence with sub-sequence.
 // 3. Internal Memory.
@@ -172,12 +172,10 @@ public:
   /**
    * InferShape must be called before Run.
    */
-  virtual void InferShape(const Scope& scope) const override {
-    alg_.InferShape(scope);
-  }
+  void InferShape(const Scope& scope) const override { alg_.InferShape(scope); }
 
-  virtual void Run(const Scope& scope,
-                   const platform::DeviceContext& dev_ctx) const override {
+  void Run(const Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override {
     alg_.Run(scope, dev_ctx);
   }
 
@@ -194,12 +192,10 @@ public:
   /**
    * InferShape must be called before Run.
    */
-  virtual void InferShape(const Scope& scope) const override {
-    alg_.InferShape(scope);
-  }
+  void InferShape(const Scope& scope) const override { alg_.InferShape(scope); }
 
-  virtual void Run(const Scope& scope,
-                   const platform::DeviceContext& dev_ctx) const override {
+  void Run(const Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override {
     alg_.Run(scope, dev_ctx);
   }
 

paddle/operators/rowwise_add_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
 #define EIGEN_USE_GPU
 #include "paddle/operators/rowwise_add_op.h"
 

paddle/operators/sgd_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
 #define EIGEN_USE_GPU
 #include "paddle/operators/sgd_op.h"
 
-REGISTER_OP_GPU_KERNEL(sgd, ops::SGDOpKernel<ops::GPUPlace, float>);
\ No newline at end of file
+REGISTER_OP_GPU_KERNEL(sgd, ops::SGDOpKernel<ops::GPUPlace, float>);

paddle/operators/sigmoid_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
 #define EIGEN_USE_GPU
 #include "paddle/operators/sigmoid_op.h"
 

paddle/operators/softmax_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
 #define EIGEN_USE_GPU
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/softmax_op.h"
 
 REGISTER_OP_GPU_KERNEL(softmax, ops::SoftmaxKernel<ops::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(softmax_grad, ops::SoftmaxGradKernel<ops::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(softmax_grad,
+                       ops::SoftmaxGradKernel<ops::GPUPlace, float>);

paddle/platform/device_context.h

@@ -40,7 +40,7 @@ class DeviceContext {
 class CPUDeviceContext : public DeviceContext {
  public:
   CPUDeviceContext();
-  CPUDeviceContext(CPUPlace);
+  explicit CPUDeviceContext(CPUPlace);
   virtual ~CPUDeviceContext() {}
 
   Eigen::DefaultDevice* eigen_device() const;
@@ -55,7 +55,7 @@ class CPUDeviceContext : public DeviceContext {
 
 class CUDADeviceContext : public DeviceContext {
  public:
-  explicit CUDADeviceContext(GPUPlace);
+  CUDADeviceContext(GPUPlace);  // NOLINT
   virtual ~CUDADeviceContext();
 
   /*! \brief  Wait for all operations completion in the stream. */
@@ -69,10 +69,10 @@ class CUDADeviceContext : public DeviceContext {
 
   // clang-format off
   /*! \brief  Return cublas handle in the device context. */
-  cublasHandle_t    cublas_handle   ();
+  cublasHandle_t    cublas_handle();
 
   /*! \brief  Return cudnn  handle in the device context. */
-  cudnnHandle_t     cudnn_handle    ();
+  cudnnHandle_t     cudnn_handle();
 
   /*! \brief  Return curand handle in the device context. */
   curandGenerator_t curand_generator();

paddle/platform/dynload/cublas.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
 #include <paddle/platform/dynload/cublas.h>
 
 namespace paddle {

paddle/platform/dynload/cudnn.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
 #include <paddle/platform/dynload/cudnn.h>
 
 namespace paddle {
@@ -25,4 +39,4 @@ CUDNN_DNN_ROUTINE_EACH_R5(DEFINE_WRAP);
 
 }  // namespace dynload
 }  // namespace platform
-}  // namespace paddle
\ No newline at end of file
+}  // namespace paddle

paddle/platform/dynload/curand.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
 #include <paddle/platform/dynload/curand.h>
 
 namespace paddle {
@@ -10,6 +24,7 @@ void *curand_dso_handle;
 #define DEFINE_WRAP(__name) DynLoad__##__name __name
 
 CURAND_RAND_ROUTINE_EACH(DEFINE_WRAP);
-}
-}
-}
\ No newline at end of file
+
+}  // namespace dynload
+}  // namespace platform
+}  // namespace paddle

paddle/platform/enforce.h

@@ -162,5 +162,50 @@ inline void throw_on_error(T e) {
     }                                                                   \
   } while (0)
 
+/*
+ * Some enforce helpers here, usage:
+ *    int a = 1;
+ *    int b = 2;
+ *    PADDLE_ENFORCE_EQ(a, b);
+ *
+ *    will raise an expression described as follows:
+ *    "enforce a == b failed, 1 != 2" with detailed stack infomation.
+ *
+ *    extra messages is also supported, for example:
+ *    PADDLE_ENFORCE(a, b, "some simple enforce failed between %d numbers", 2)
+ */
+
+#define PADDLE_ENFORCE_EQ(__VAL0, __VAL1, ...) \
+  __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, ==, !=, __VA_ARGS__)
+#define PADDLE_ENFORCE_NE(__VAL0, __VAL1, ...) \
+  __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, !=, ==, __VA_ARGS__)
+#define PADDLE_ENFORCE_GT(__VAL0, __VAL1, ...) \
+  __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, >, <=, __VA_ARGS__)
+#define PADDLE_ENFORCE_GE(__VAL0, __VAL1, ...) \
+  __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, >=, <, __VA_ARGS__)
+#define PADDLE_ENFORCE_LT(__VAL0, __VAL1, ...) \
+  __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, <, >=, __VA_ARGS__)
+#define PADDLE_ENFORCE_LE(__VAL0, __VAL1, ...) \
+  __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, <=, >, __VA_ARGS__)
+
+// if two values have different data types, choose a compatible type for them.
+template <typename T1, typename T2>
+struct CompatibleType {
+  static const bool t1_to_t2 = std::is_convertible<T1, T2>::value;
+  typedef typename std::conditional<t1_to_t2, T2, T1>::type type;
+};
+
+#define __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, __CMP, __INV_CMP, ...)        \
+  PADDLE_ENFORCE(__COMPATIBLE_TYPE(__VAL0, __VAL1, __VAL0)                    \
+                     __CMP __COMPATIBLE_TYPE(__VAL0, __VAL1, __VAL1),         \
+                 "enforce %s " #__CMP " %s failed, %s " #__INV_CMP " %s\n%s", \
+                 #__VAL0, #__VAL1, std::to_string(__VAL0),                    \
+                 std::to_string(__VAL1),                                      \
+                 paddle::string::Sprintf("" __VA_ARGS__));
+
+#define __COMPATIBLE_TYPE(__VAL0, __VAL1, __VAL)              \
+  typename paddle::platform::CompatibleType<decltype(__VAL0), \
+                                            decltype(__VAL1)>::type(__VAL)
+
 }  // namespace platform
 }  // namespace paddle

paddle/platform/enforce_test.cc

@@ -34,3 +34,165 @@ TEST(ENFORCE, FAILED) {
   }
   ASSERT_TRUE(in_catch);
 }
+
+TEST(ENFORCE, NO_ARG_OK) {
+  int a = 2;
+  int b = 2;
+  PADDLE_ENFORCE_EQ(a, b);
+  // test enforce with extra message.
+  PADDLE_ENFORCE_EQ(a, b, "some thing wrong %s", "info");
+}
+
+TEST(ENFORCE_EQ, NO_EXTRA_MSG_FAIL) {
+  int a = 2;
+  bool in_catch = false;
+
+  try {
+    PADDLE_ENFORCE_EQ(a, 1 + 3);
+
+  } catch (paddle::platform::EnforceNotMet error) {
+    in_catch = true;
+    const std::string msg = "enforce a == 1 + 3 failed, 2 != 4";
+    const char* what = error.what();
+    for (size_t i = 0; i < msg.length(); ++i) {
+      ASSERT_EQ(what[i], msg[i]);
+    }
+  }
+
+  ASSERT_TRUE(in_catch);
+}
+
+TEST(ENFORCE_EQ, EXTRA_MSG_FAIL) {
+  int a = 2;
+  bool in_catch = false;
+
+  try {
+    PADDLE_ENFORCE_EQ(a, 1 + 3, "%s size not match", "their");
+
+  } catch (paddle::platform::EnforceNotMet error) {
+    in_catch = true;
+    const std::string msg =
+        "enforce a == 1 + 3 failed, 2 != 4\ntheir size not match";
+    const char* what = error.what();
+    for (size_t i = 0; i < msg.length(); ++i) {
+      ASSERT_EQ(what[i], msg[i]);
+    }
+  }
+
+  ASSERT_TRUE(in_catch);
+}
+
+TEST(ENFORCE_NE, OK) {
+  PADDLE_ENFORCE_NE(1, 2);
+  PADDLE_ENFORCE_NE(1.0, 2UL);
+}
+TEST(ENFORCE_NE, FAIL) {
+  bool in_catch = false;
+
+  try {
+    // 2UL here to check data type compatible
+    PADDLE_ENFORCE_NE(1.0, 1UL);
+
+  } catch (paddle::platform::EnforceNotMet error) {
+    in_catch = true;
+    const std::string msg = "enforce 1.0 != 1UL failed, 1.000000 == 1";
+    const char* what = error.what();
+    for (size_t i = 0; i < msg.length(); ++i) {
+      ASSERT_EQ(what[i], msg[i]);
+    }
+  }
+
+  ASSERT_TRUE(in_catch);
+}
+
+TEST(ENFORCE_GT, OK) { PADDLE_ENFORCE_GT(2, 1); }
+TEST(ENFORCE_GT, FAIL) {
+  bool in_catch = false;
+
+  try {
+    // 2UL here to check data type compatible
+    PADDLE_ENFORCE_GT(1, 2UL);
+
+  } catch (paddle::platform::EnforceNotMet error) {
+    in_catch = true;
+    const std::string msg = "enforce 1 > 2UL failed, 1 <= 2";
+    const char* what = error.what();
+    for (size_t i = 0; i < msg.length(); ++i) {
+      ASSERT_EQ(what[i], msg[i]);
+    }
+  }
+
+  ASSERT_TRUE(in_catch);
+}
+
+TEST(ENFORCE_GE, OK) {
+  PADDLE_ENFORCE_GE(2, 2UL);
+  PADDLE_ENFORCE_GE(3, 2UL);
+  PADDLE_ENFORCE_GE(3, 2);
+  PADDLE_ENFORCE_GE(3.21, 2UL);
+}
+TEST(ENFORCE_GE, FAIL) {
+  bool in_catch = false;
+
+  try {
+    PADDLE_ENFORCE_GE(1, 2UL);
+
+  } catch (paddle::platform::EnforceNotMet error) {
+    in_catch = true;
+    const std::string msg = "enforce 1 >= 2UL failed, 1 < 2";
+    const char* what = error.what();
+    for (size_t i = 0; i < msg.length(); ++i) {
+      ASSERT_EQ(what[i], msg[i]);
+    }
+  }
+
+  ASSERT_TRUE(in_catch);
+}
+
+TEST(ENFORCE_LE, OK) {
+  PADDLE_ENFORCE_LE(1, 1);
+  PADDLE_ENFORCE_LE(1, 1UL);
+  PADDLE_ENFORCE_LE(2, 3UL);
+  PADDLE_ENFORCE_LE(2UL, 3);
+  PADDLE_ENFORCE_LE(2UL, 3.2);
+}
+TEST(ENFORCE_LE, FAIL) {
+  bool in_catch = false;
+
+  try {
+    PADDLE_ENFORCE_GT(1, 2UL);
+
+  } catch (paddle::platform::EnforceNotMet error) {
+    in_catch = true;
+    const std::string msg = "enforce 1 > 2UL failed, 1 <= 2";
+    const char* what = error.what();
+    for (size_t i = 0; i < msg.length(); ++i) {
+      ASSERT_EQ(what[i], msg[i]);
+    }
+  }
+
+  ASSERT_TRUE(in_catch);
+}
+
+TEST(ENFORCE_LT, OK) {
+  PADDLE_ENFORCE_LT(3, 10);
+  PADDLE_ENFORCE_LT(2, 3UL);
+  PADDLE_ENFORCE_LT(2UL, 3);
+}
+TEST(ENFORCE_LT, FAIL) {
+  bool in_catch = false;
+
+  try {
+    PADDLE_ENFORCE_LT(1UL, 0.12);
+
+  } catch (paddle::platform::EnforceNotMet error) {
+    in_catch = true;
+    const std::string msg = "enforce 1UL < 0.12 failed, 1 >= 0.12";
+    const char* what = error.what();
+    for (size_t i = 0; i < msg.length(); ++i) {
+      ASSERT_EQ(what[i], msg[i]);
+    }
+  }
+
+  ASSERT_TRUE(in_catch);
+}

paddle/platform/place.h

@@ -32,7 +32,7 @@ struct CPUPlace {
 
 struct GPUPlace {
   GPUPlace() : GPUPlace(0) {}
-  GPUPlace(int d) : device(d) {}
+  GPUPlace(int d) : device(d) {}  // NOLINT
 
   // needed for variant equality comparison
   inline bool operator==(const GPUPlace &o) const { return device == o.device; }

paddle/string/piece.h

@@ -39,8 +39,8 @@ public:
   // size_ is 0.
   Piece();
   Piece(const char* d, size_t n);
-  Piece(const char* d);
-  Piece(const std::string& s);
+  Piece(const char* d);         // NOLINT
+  Piece(const std::string& s);  // NOLINT
 
   const char* data() const { return data_; }
   size_t len() const { return size_; }

python/paddle/v2/dataset/cifar.py

@@ -133,7 +133,7 @@ def convert(path):
     """
     Converts dataset to recordio format
     """
-    paddle.v2.dataset.common.convert(path, train100(), 10, "cifar_train100")
-    paddle.v2.dataset.common.convert(path, test100(), 10, "cifar_test100")
-    paddle.v2.dataset.common.convert(path, train10(), 10, "cifar_train10")
-    paddle.v2.dataset.common.convert(path, test10(), 10, "cifar_test10")
+    paddle.v2.dataset.common.convert(path, train100(), 1000, "cifar_train100")
+    paddle.v2.dataset.common.convert(path, test100(), 1000, "cifar_test100")
+    paddle.v2.dataset.common.convert(path, train10(), 1000, "cifar_train10")
+    paddle.v2.dataset.common.convert(path, test10(), 1000, "cifar_test10")

python/paddle/v2/dataset/common.py

@@ -32,17 +32,22 @@ __all__ = [
 
 DATA_HOME = os.path.expanduser('~/.cache/paddle/dataset')
 
+
 # When running unit tests, there could be multiple processes that
 # trying to create DATA_HOME directory simultaneously, so we cannot
 # use a if condition to check for the existence of the directory;
 # instead, we use the filesystem as the synchronization mechanism by
 # catching returned errors.
-try:
-    os.makedirs(DATA_HOME)
-except OSError as exc:
-    if exc.errno != errno.EEXIST:
-        raise
-    pass
+def must_mkdirs(path):
+    try:
+        os.makedirs(DATA_HOME)
+    except OSError as exc:
+        if exc.errno != errno.EEXIST:
+            raise
+        pass
+
+
+must_mkdirs(DATA_HOME)
 
 
 def md5file(fname):
@@ -93,6 +98,19 @@ def fetch_all():
                 "fetch")()
 
 
+def fetch_all_recordio(path):
+    for module_name in filter(lambda x: not x.startswith("__"),
+                              dir(paddle.v2.dataset)):
+        if "convert" in dir(
+                importlib.import_module("paddle.v2.dataset.%s" % module_name)) and \
+                not module_name == "common":
+            ds_path = os.path.join(path, module_name)
+            must_mkdirs(ds_path)
+            getattr(
+                importlib.import_module("paddle.v2.dataset.%s" % module_name),
+                "convert")(ds_path)
+
+
 def split(reader, line_count, suffix="%05d.pickle", dumper=cPickle.dump):
     """
     you can call the function as:

python/paddle/v2/dataset/conll05.py

@@ -233,5 +233,5 @@ def convert(path):
     """
     Converts dataset to recordio format
     """
-    paddle.v2.dataset.common.convert(path, test(), 10, "conl105_train")
-    paddle.v2.dataset.common.convert(path, test(), 10, "conl105_test")
+    paddle.v2.dataset.common.convert(path, test(), 1000, "conl105_train")
+    paddle.v2.dataset.common.convert(path, test(), 1000, "conl105_test")

python/paddle/v2/dataset/imdb.py

@@ -173,5 +173,5 @@ def convert(path):
     Converts dataset to recordio format
     """
     w = word_dict()
-    paddle.v2.dataset.common.convert(path, lambda: train(w), 10, "imdb_train")
-    paddle.v2.dataset.common.convert(path, lambda: test(w), 10, "imdb_test")
+    paddle.v2.dataset.common.convert(path, lambda: train(w), 1000, "imdb_train")
+    paddle.v2.dataset.common.convert(path, lambda: test(w), 1000, "imdb_test")

python/paddle/v2/dataset/imikolov.py

@@ -155,6 +155,7 @@ def convert(path):
     N = 5
     word_dict = build_dict()
     paddle.v2.dataset.common.convert(path,
-                                     train(word_dict, N), 10, "imikolov_train")
+                                     train(word_dict, N), 1000,
+                                     "imikolov_train")
     paddle.v2.dataset.common.convert(path,
-                                     test(word_dict, N), 10, "imikolov_test")
+                                     test(word_dict, N), 1000, "imikolov_test")

python/paddle/v2/dataset/mnist.py

@@ -119,5 +119,5 @@ def convert(path):
     """
     Converts dataset to recordio format
     """
-    paddle.v2.dataset.common.convert(path, train(), 10, "minist_train")
-    paddle.v2.dataset.common.convert(path, test(), 10, "minist_test")
+    paddle.v2.dataset.common.convert(path, train(), 1000, "minist_train")
+    paddle.v2.dataset.common.convert(path, test(), 1000, "minist_test")

python/paddle/v2/dataset/movielens.py

@@ -254,8 +254,8 @@ def convert(path):
     """
     Converts dataset to recordio format
     """
-    paddle.v2.dataset.common.convert(path, train(), 10, "movielens_train")
-    paddle.v2.dataset.common.convert(path, test(), 10, "movielens_test")
+    paddle.v2.dataset.common.convert(path, train(), 1000, "movielens_train")
+    paddle.v2.dataset.common.convert(path, test(), 1000, "movielens_test")
 
 
 if __name__ == '__main__':

python/paddle/v2/dataset/sentiment.py

@@ -137,5 +137,5 @@ def convert(path):
     """
     Converts dataset to recordio format
     """
-    paddle.v2.dataset.common.convert(path, train, 10, "sentiment_train")
-    paddle.v2.dataset.common.convert(path, test, 10, "sentiment_test")
+    paddle.v2.dataset.common.convert(path, train, 1000, "sentiment_train")
+    paddle.v2.dataset.common.convert(path, test, 1000, "sentiment_test")

python/paddle/v2/dataset/uci_housing.py

@@ -119,5 +119,5 @@ def convert(path):
     """
     Converts dataset to recordio format
     """
-    paddle.v2.dataset.common.convert(path, train(), 10, "uci_housing_train")
-    paddle.v2.dataset.common.convert(path, test(), 10, "uci_houseing_test")
+    paddle.v2.dataset.common.convert(path, train(), 1000, "uci_housing_train")
+    paddle.v2.dataset.common.convert(path, test(), 1000, "uci_houseing_test")

python/paddle/v2/dataset/wmt14.py

@@ -169,5 +169,6 @@ def convert(path):
     Converts dataset to recordio format
     """
     dict_size = 30000
-    paddle.v2.dataset.common.convert(path, train(dict_size), 10, "wmt14_train")
-    paddle.v2.dataset.common.convert(path, test(dict_size), 10, "wmt14_test")
+    paddle.v2.dataset.common.convert(path,
+                                     train(dict_size), 1000, "wmt14_train")
+    paddle.v2.dataset.common.convert(path, test(dict_size), 1000, "wmt14_test")

python/paddle/v2/framework/tests/CMakeLists.txt

@@ -13,4 +13,5 @@ add_python_test(test_framework
     test_sigmoid_op.py
     test_softmax_op.py
     test_rowwise_add_op.py
-    test_network.py)
+    test_network.py
+    gradient_checker.py)

python/paddle/v2/framework/tests/gradient_checker.py

+import paddle.v2.framework.core as core
+from paddle.v2.framework.create_op_creation_methods import op_creations
+import numpy
+import unittest
+
+__all__ = ['get_numeric_gradient']
+
+
+def get_numeric_gradient(op,
+                         input_values,
+                         output_name,
+                         input_to_check,
+                         delta=1e-2,
+                         local_scope=None):
+    """
+    Get Numeric Gradient for an operator's input.
+    
+    :param op: C++ operator instance, could be an network 
+    :param input_values: The input variables. Should be an dictionary, key is 
+    variable name. Value is numpy array.
+    :param output_name: The final output variable name. 
+    :param input_to_check: The input variable need to get gradient.
+    :param delta: The perturbation value for numeric gradient method. The 
+    smaller delta is, the more accurate result will get. But if that delta is
+     too small, it could occur numerical stability problem.
+    :param local_scope: The local scope used for get_numeric_gradient.
+    :return: The gradient array in numpy format.
+    """
+    if local_scope is None:
+        local_scope = core.Scope()
+
+    # Create all input variable in local_scope
+    for var_name in input_values:
+        var = local_scope.new_var(var_name)
+        tensor = var.get_tensor()
+        tensor.set_dims(input_values[var_name].shape)
+        tensor.alloc_float(core.CPUPlace())
+        tensor.set(input_values[var_name], core.CPUPlace())
+
+    # Create all output variable in local_scope
+    for output in op.outputs():
+        if local_scope.find_var(output) is None:
+            local_scope.new_var(output).get_tensor()
+
+    op.infer_shape(local_scope)
+
+    # allocate output memory
+    for output in op.outputs():
+        local_scope.find_var(output).get_tensor().alloc_float(core.CPUPlace())
+
+    # TODO(yuyang18): Only CPU is support now.
+    cpu_ctx = core.DeviceContext.create(core.CPUPlace())
+
+    def get_output():
+        op.run(local_scope, cpu_ctx)
+        return numpy.array(local_scope.find_var(output_name).get_tensor()).sum()
+
+    def product(dim):
+        return reduce(lambda a, b: a * b, dim, 1)
+
+    tensor_to_check = local_scope.find_var(input_to_check).get_tensor()
+    tensor_size = product(tensor_to_check.get_dims())
+    gradient_flat = numpy.zeros(shape=(tensor_size, ), dtype='float32')
+    for i in xrange(tensor_size):
+        origin = tensor_to_check.get_float_element(i)
+        x_pos = origin + delta
+        tensor_to_check.set_float_element(i, x_pos)
+        y_pos = get_output()
+
+        x_neg = origin - delta
+        tensor_to_check.set_float_element(i, x_neg)
+        y_neg = get_output()
+
+        tensor_to_check.set_float_element(i, origin)  # restore old value
+        gradient_flat[i] = (y_pos - y_neg) / delta / 2
+    return gradient_flat.reshape(tensor_to_check.get_dims())
+
+
+if __name__ == '__main__':
+
+    class GetNumericGradientTest(unittest.TestCase):
+        def test_add_op(self):
+            add_op = op_creations.add_two(X="X", Y="Y", Out="Z")
+            x = numpy.random.random((10, 1)).astype("float32")
+            y = numpy.random.random((10, 1)).astype("float32")
+
+            arr = get_numeric_gradient(add_op, {'X': x, "Y": y}, 'Z', 'X')
+            self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-2)
+
+    unittest.main()