Merge branch 'develop' of github.com:baidu/Paddle into feature/fast_python_unittest

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/cpplint.cmake

@@ -56,11 +56,14 @@ macro(add_style_check_target TARGET_NAME)
                 # cpplint code style
                 get_filename_component(base_filename ${filename} NAME)
                 set(CUR_GEN ${CMAKE_CURRENT_BINARY_DIR}/${base_filename}.cpplint)
-                add_custom_command(TARGET ${TARGET_NAME} PRE_BUILD
+                add_custom_command(OUTPUT ${CUR_GEN} PRE_BUILD
                     COMMAND "${PYTHON_EXECUTABLE}" "${PROJ_ROOT}/paddle/scripts/cpplint.py"
                             "--filter=${STYLE_FILTER}"
                             "--write-success=${CUR_GEN}" ${filename}
+                    DEPENDS ${filename} ${PROJ_ROOT}/paddle/scripts/cpplint.py
                     WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
+                add_custom_target(${base_filename}.cpplint DEPENDS ${CUR_GEN})
+                add_dependencies(${TARGET_NAME} ${base_filename}.cpplint)
             endif()
         endforeach()
     endif()

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})

cmake/util.cmake

@@ -118,7 +118,6 @@ endfunction()
 macro(add_unittest_without_exec TARGET_NAME)
     add_executable(${TARGET_NAME} ${ARGN})
     link_paddle_test(${TARGET_NAME})
-    add_style_check_target(${TARGET_NAME} ${ARGN})
 endmacro()
 
 # add_unittest

paddle/CMakeLists.txt

@@ -15,7 +15,6 @@ if(Boost_FOUND)
   add_subdirectory(platform)
   add_subdirectory(framework)
   add_subdirectory(operators)
-  add_subdirectory(pybind)
 endif()
 
 if(WITH_C_API)

paddle/framework/CMakeLists.txt

@@ -12,13 +12,15 @@ cc_test(variable_test SRCS variable_test.cc)
 cc_library(scope SRCS scope.cc)
 cc_test(scope_test SRCS scope_test.cc DEPS scope)
 
-proto_library(attr_type SRCS attr_type.proto)
-proto_library(op_proto SRCS op_proto.proto DEPS attr_type)
-proto_library(op_desc SRCS op_desc.proto DEPS attr_type)
+proto_library(attribute_proto SRCS attribute.proto)
+proto_library(op_proto SRCS op_proto.proto DEPS attribute_proto)
+proto_library(op_desc SRCS op_desc.proto DEPS attribute_proto)
 cc_test(op_proto_test SRCS op_proto_test.cc DEPS op_proto protobuf)
 cc_test(op_desc_test SRCS op_desc_test.cc DEPS op_desc protobuf)
 
-cc_library(operator SRCS operator.cc DEPS op_desc device_context tensor scope)
+cc_library(attribute SRCS attribute.cc DEPS op_desc op_proto)
+
+cc_library(operator SRCS operator.cc DEPS op_desc device_context tensor scope attribute)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
 
 cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS op_proto operator)
@@ -26,13 +28,19 @@ cc_library(op_registry SRCS op_registry.cc DEPS op_desc grad_op_builder)
 cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
 cc_test(grad_op_builder_test SRCS grad_op_builder_test.cc DEPS grad_op_builder op_registry add_op)
 
-py_proto_compile(framework_py_proto SRCS attr_type.proto op_proto.proto op_desc.proto)
+py_proto_compile(framework_py_proto SRCS attribute.proto op_proto.proto op_desc.proto)
 # Generate an empty __init__.py to make framework_py_proto as a valid python module.
 add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
 add_dependencies(framework_py_proto framework_py_proto_init)
 
-cc_library(net SRCS net.cc DEPS op_registry)
-cc_test(net_op_test SRCS net_op_test.cc DEPS net)
-
-cc_library(backward SRCS backward.cc DEPS net)
+cc_library(backward SRCS backward.cc DEPS net_op)
 cc_test(backward_test SRCS backward_test.cc DEPS backward)
+cc_library(paddle_pybind SHARED
+    SRCS pybind.cc
+    DEPS pybind python backward
+	fc_op
+	sgd_op
+	add_op
+	mean_op
+	cross_entropy_op
+	recurrent_op)

paddle/framework/attribute.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/framework/attribute.h"
+
+#include <vector>
+
+namespace paddle {
+namespace framework {
+
+template <>
+AttrType AttrTypeID<int>() {
+  return INT;
+}
+template <>
+AttrType AttrTypeID<float>() {
+  return FLOAT;
+}
+template <>
+AttrType AttrTypeID<std::string>() {
+  return STRING;
+}
+template <>
+AttrType AttrTypeID<std::vector<int>>() {
+  return INTS;
+}
+template <>
+AttrType AttrTypeID<std::vector<float>>() {
+  return FLOATS;
+}
+template <>
+AttrType AttrTypeID<std::vector<std::string>>() {
+  return STRINGS;
+}
+
+Attribute GetAttrValue(const AttrDesc& attr_desc) {
+  switch (attr_desc.type()) {
+    case paddle::framework::AttrType::INT: {
+      return attr_desc.i();
+    }
+    case paddle::framework::AttrType::FLOAT: {
+      return attr_desc.f();
+    }
+    case paddle::framework::AttrType::STRING: {
+      return attr_desc.s();
+    }
+    case paddle::framework::AttrType::INTS: {
+      std::vector<int> val(attr_desc.ints_size());
+      for (int i = 0; i < attr_desc.ints_size(); ++i) {
+        val[i] = attr_desc.ints(i);
+      }
+      return val;
+    }
+    case paddle::framework::AttrType::FLOATS: {
+      std::vector<float> val(attr_desc.floats_size());
+      for (int i = 0; i < attr_desc.floats_size(); ++i) {
+        val[i] = attr_desc.floats(i);
+      }
+      return val;
+    }
+    case paddle::framework::AttrType::STRINGS: {
+      std::vector<std::string> val(attr_desc.strings_size());
+      for (int i = 0; i < attr_desc.strings_size(); ++i) {
+        val[i] = attr_desc.strings(i);
+      }
+      return val;
+    }
+  }
+  PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !");
+  return boost::blank();
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/attr_checker.h → paddle/framework/attribute.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 <boost/variant.hpp>
@@ -6,6 +20,9 @@
 #include <unordered_map>
 #include <unordered_set>
 #include <vector>
+
+#include "paddle/framework/attribute.pb.h"
+#include "paddle/framework/op_desc.pb.h"
 #include "paddle/platform/enforce.h"
 
 namespace paddle {
@@ -14,13 +31,19 @@ namespace framework {
 typedef boost::variant<boost::blank, int, float, std::string, std::vector<int>,
                        std::vector<float>, std::vector<std::string>>
     Attribute;
+
 typedef std::unordered_map<std::string, Attribute> AttributeMap;
 
+template <typename T>
+AttrType AttrTypeID();
+
+Attribute GetAttrValue(const AttrDesc& attr_desc);
+
 // check whether a value(attribute) fit a certain limit
 template <typename T>
 class LargerThanChecker {
  public:
-  LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
+  explicit LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
   void operator()(T& value) const {
     PADDLE_ENFORCE(value > lower_bound_, "larger_than check fail");
   }
@@ -35,7 +58,8 @@ class LargerThanChecker {
 template <typename T>
 class DefaultValueSetter {
  public:
-  DefaultValueSetter(T default_value) : default_value_(default_value) {}
+  explicit DefaultValueSetter(T default_value)
+      : default_value_(default_value) {}
   void operator()(T& value) const { value = default_value_; }
 
  private:
@@ -78,7 +102,8 @@ class TypedAttrChecker {
   typedef std::function<void(T&)> ValueChecker;
 
  public:
-  TypedAttrChecker(const std::string& attr_name) : attr_name_(attr_name) {}
+  explicit TypedAttrChecker(const std::string& attr_name)
+      : attr_name_(attr_name) {}
 
   TypedAttrChecker& InEnum(const std::unordered_set<T>& range) {
     value_checkers_.push_back(EnumInContainer<T>(range));

paddle/framework/backward.cc

@@ -14,8 +14,8 @@
 
 #include "paddle/framework/backward.h"
 #include <list>
-#include "paddle/framework/net.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/net_op.h"
 
 namespace paddle {
 namespace framework {
@@ -32,7 +32,7 @@ static bool AllInSet(const std::vector<std::string>& names,
 }
 
 static std::shared_ptr<OperatorBase> NOP() {
-  auto net_op = std::make_shared<NetOp>();
+  auto net_op = std::make_shared<operators::NetOp>();
   net_op->type_ = "@NOP@";
   net_op->CompleteAddOp();
   return net_op;
@@ -42,9 +42,9 @@ static std::shared_ptr<OperatorBase> NOP() {
 //
 //  no_grad_names the gradient variable names without gradient calculating.
 //
-//  uniq_id is a unique index used inside recursively calling BackwardRecursive.
-//  use `uid = uniq_id++;` to get the unique index, and pass `uniq_id` through
-//  recursive calling.
+//  uniq_id is a unique index used inside recursively calling
+//  BackwardRecursive. use `uid = uniq_id++;` to get the unique index, and
+//  pass `uniq_id` through recursive calling.
 //
 //  returns The backward operator. For simple situation, it is a simple
 //  operator. For complex situation, it is a NetOp.
@@ -59,32 +59,30 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
   //  If all input gradients of forwarding operator do not need to calculate,
   //  just return an NOP. Not return null ptr because NOP does not take
   //  too much time for calculation, but it is useful for simplifying logic.
-  if (AllInSet(forwardOp.inputs_, OperatorBase::GRAD_VAR_SUFFIX(),
-               no_grad_names)) {
+  if (AllInSet(forwardOp.inputs_, kGradVarSuffix, no_grad_names)) {
     return NOP();
   }
 
-  //  All output gradients of forwarding operator do not need to calculate. Then
-  //  all input gradients cannot be computed at all, and we put them into
+  //  All output gradients of forwarding operator do not need to calculate.
+  //  Then all input gradients cannot be computed at all, and we put them into
   //  `no_grad_names` set. Return an NOP.
-  if (AllInSet(forwardOp.outputs_, OperatorBase::GRAD_VAR_SUFFIX(),
-               no_grad_names)) {
+  if (AllInSet(forwardOp.outputs_, kGradVarSuffix, no_grad_names)) {
     for (auto& name : forwardOp.inputs_) {
       // Mark all input is not need
-      no_grad_names.insert(name + OperatorBase::GRAD_VAR_SUFFIX());
+      no_grad_names.insert(name + kGradVarSuffix);
     }
     return NOP();
   }
 
   // Returned gradient network
-  auto net = std::make_shared<NetOp>();
+  auto net = std::make_shared<operators::NetOp>();
 
   if (forwardOp.IsNetOp()) {
     // Because forwardOp is a net op, it can static_cast.
-    auto& forwardNet = static_cast<const NetOp&>(forwardOp);
+    auto& forwardNet = static_cast<const operators::NetOp&>(forwardOp);
 
-    // Map from output gradient variable name to operator's indices in backward
-    // net. That operator generates that variable.
+    // Map from output gradient variable name to operator's indices in
+    // backward net. That operator generates that variable.
     std::unordered_map<std::string, std::vector<size_t>> dup_output_ops;
 
     size_t local_op_id = 0;
@@ -134,9 +132,9 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
     std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
     for (std::string& grad_input : grad_op->inputs_) {
       if (no_grad_names.count(grad_input)) {
-        std::string prefix = grad_input.substr(
-            0, grad_input.size() - OperatorBase::GRAD_VAR_SUFFIX().size());
-        grad_input = prefix + OperatorBase::ZERO_VAR_SUFFIX();
+        std::string prefix =
+            grad_input.substr(0, grad_input.size() - kGradVarSuffix.size());
+        grad_input = prefix + kZeroVarSuffix;
 
         // If part of input gradient of that operator is not calculated, fill
         // zero variables to that input gradient.
@@ -147,7 +145,7 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
 
     for (std::string& grad_output : grad_op->outputs_) {
       if (no_grad_names.count(grad_output)) {
-        grad_output = OperatorBase::EMPTY_VAR_NAME();
+        grad_output = kEmptyVarName;
       }
     }
 
@@ -168,14 +166,14 @@ std::shared_ptr<OperatorBase> Backward(
   std::unordered_set<std::string> no_grad_names;
   no_grad_names.reserve(no_grad_vars.size());
 
-  no_grad_names.insert(OperatorBase::EMPTY_VAR_NAME() +
-                       OperatorBase::GRAD_VAR_SUFFIX());
+  no_grad_names.insert(kEmptyVarName + kGradVarSuffix);
 
   for (auto& name : no_grad_vars) {
-    no_grad_names.insert(name + OperatorBase::GRAD_VAR_SUFFIX());
+    no_grad_names.insert(name + kGradVarSuffix);
   }
   size_t uid = 0;
   return BackwardRecursive(forwardOp, no_grad_names, uid);
 }
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/backward_test.cc

@@ -15,8 +15,9 @@
 #include "paddle/framework/backward.h"
 
 #include <gtest/gtest.h>
-#include "paddle/framework/net.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/net_op.h"
+#include "paddle/operators/type_alias.h"
 
 namespace paddle {
 namespace framework {
@@ -70,21 +71,21 @@ class NoGradOpMaker : public OpProtoAndCheckerMaker {
   }
 };
 
-class FcOp : public NetOp {
+class FcOp : public ops::NetOp {
  public:
   void Init() override {
     AddOp(OpRegistry::CreateOp("mul", {Input("X"), Input("W")},
                                {Output("mul_result")}, {}));
     auto b_name = Input("b");
     std::string before_act = "mul_result";
-    if (b_name != EMPTY_VAR_NAME()) {
+    if (b_name != kEmptyVarName) {
       AddOp(OpRegistry::CreateOp("rowwise_add", {Output("mul_result"), b_name},
                                  {Output("add_result")}, {}));
       before_act = "add_result";
     } else {
       auto out_varname = Output("add_result");
-      if (out_varname != EMPTY_VAR_NAME()) {
-        this->Rename(out_varname, EMPTY_VAR_NAME());
+      if (out_varname != kEmptyVarName) {
+        this->Rename(out_varname, kEmptyVarName);
       }
     }
 
@@ -161,14 +162,13 @@ TEST(Backward, simple_op_grad) {
   auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
   ASSERT_NE(fwd, nullptr);
   auto gop = f::OpRegistry::CreateGradOp(*fwd);
-  ASSERT_EQ(1UL, gop->inputs_.size());
-  ASSERT_EQ("Out" + f::OperatorBase::GRAD_VAR_SUFFIX(), gop->inputs_[0]);
+  ASSERT_EQ(4UL, gop->inputs_.size());
+  ASSERT_EQ(f::kEmptyVarName, gop->inputs_[0]);
   ASSERT_EQ("rowwise_add_grad", gop->type_);
-  ASSERT_EQ("X" + f::OperatorBase::GRAD_VAR_SUFFIX(), gop->outputs_[0]);
-  ASSERT_EQ("b" + f::OperatorBase::GRAD_VAR_SUFFIX(), gop->outputs_[1]);
+  ASSERT_EQ("X" + f::kGradVarSuffix, gop->outputs_[0]);
+  ASSERT_EQ("b" + f::kGradVarSuffix, gop->outputs_[1]);
 
-  ASSERT_EQ("X" + f::OperatorBase::GRAD_VAR_SUFFIX(),
-            gop->Output("X" + f::OperatorBase::GRAD_VAR_SUFFIX()));
+  ASSERT_EQ("X" + f::kGradVarSuffix, gop->Output("X" + f::kGradVarSuffix));
 }
 
 TEST(Backward, simple_op_not_need_grad) {
@@ -176,13 +176,14 @@ TEST(Backward, simple_op_not_need_grad) {
   ASSERT_NE(fwd, nullptr);
   auto gop = f::Backward(*fwd, {"X"});
   ASSERT_EQ(std::find(gop->outputs_.begin(), gop->outputs_.end(),
-                      "X" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+                      "X" + f::kGradVarSuffix),
             gop->outputs_.end());
 
   auto no_input_gop = f::Backward(*fwd, {"X", "b"});
   ASSERT_NE(no_input_gop, nullptr);
   ASSERT_TRUE(no_input_gop->IsNetOp());
-  ASSERT_EQ(0UL, std::static_pointer_cast<f::NetOp>(no_input_gop)->ops_.size());
+  ASSERT_EQ(0UL,
+            std::static_pointer_cast<ops::NetOp>(no_input_gop)->ops_.size());
 }
 
 TEST(Backward, net_fc_backward_normal) {
@@ -191,7 +192,7 @@ TEST(Backward, net_fc_backward_normal) {
   ASSERT_NE(fwd, nullptr);
   std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
   ASSERT_TRUE(gop->IsNetOp());
-  auto net = static_cast<f::NetOp *>(gop.get());
+  auto net = static_cast<ops::NetOp *>(gop.get());
 
   ASSERT_NO_THROW(net->DebugString());
 
@@ -208,13 +209,13 @@ TEST(Backward, net_fc_backward_normal) {
 }
 
 TEST(Backward, net_fc_backward_not_have_b) {
-  std::shared_ptr<f::OperatorBase> fwd = f::OpRegistry::CreateOp(
-      "fc", {"X", "w", f::OperatorBase::EMPTY_VAR_NAME()},
-      {"mul_result", "add_result", "tmp"}, {});
+  std::shared_ptr<f::OperatorBase> fwd =
+      f::OpRegistry::CreateOp("fc", {"X", "w", f::kEmptyVarName},
+                              {"mul_result", "add_result", "tmp"}, {});
   ASSERT_NE(fwd, nullptr);
   std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
   ASSERT_TRUE(gop->IsNetOp());
-  auto net = static_cast<f::NetOp *>(gop.get());
+  auto net = static_cast<ops::NetOp *>(gop.get());
 
   ASSERT_NO_THROW(net->DebugString());
 
@@ -228,7 +229,7 @@ TEST(Backward, net_fc_backward_not_have_b) {
 }
 
 TEST(Backward, net_input_of_network_not_need_grad) {
-  f::NetOp net;
+  ops::NetOp net;
   net.AddOp(f::OpRegistry::CreateOp("fc", {"X", "W1", "b1"},
                                     {"mul_tmp_0", "add_tmp_0", "hidden0"}, {}));
   net.AddOp(f::OpRegistry::CreateOp("fc", {"hidden0", "W2", "b2"},
@@ -236,39 +237,36 @@ TEST(Backward, net_input_of_network_not_need_grad) {
   net.CompleteAddOp();
   auto bwd = Backward(net, {"X"});  // X@GRAD is not need.
   ASSERT_TRUE(bwd->IsNetOp());
-  auto bwd_net = static_cast<f::NetOp *>(bwd.get());
+  auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
 
   std::unordered_set<std::string> all_output = std::unordered_set<std::string>(
       bwd_net->outputs_.begin(), bwd_net->outputs_.end());
-  all_output.erase(f::OperatorBase::EMPTY_VAR_NAME());
+  all_output.erase(f::kEmptyVarName);
 
   for (auto &out : {"W1", "b1", "hidden0", "W2", "b2"}) {
-    ASSERT_NE(all_output.find(out + f::OperatorBase::GRAD_VAR_SUFFIX()),
-              all_output.end());
+    ASSERT_NE(all_output.find(out + f::kGradVarSuffix), all_output.end());
   }
 
   // Not Generated X
-  ASSERT_EQ(all_output.find("X" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-            all_output.end());
+  ASSERT_EQ(all_output.find("X" + f::kGradVarSuffix), all_output.end());
 
   ASSERT_EQ(2UL, bwd_net->ops_.size());
   ASSERT_TRUE(bwd_net->ops_[1]->IsNetOp());
-  auto first_fc_grad = static_cast<f::NetOp *>(bwd_net->ops_[1].get());
+  auto first_fc_grad = static_cast<ops::NetOp *>(bwd_net->ops_[1].get());
   ASSERT_EQ(3UL, first_fc_grad->ops_.size());
-  ASSERT_EQ(
-      f::OperatorBase::EMPTY_VAR_NAME(),
-      first_fc_grad->ops_[2]->Output("A" + f::OperatorBase::GRAD_VAR_SUFFIX()));
+  ASSERT_EQ(f::kEmptyVarName,
+            first_fc_grad->ops_[2]->Output("A" + f::kGradVarSuffix));
 }
 
 TEST(Backward, net_shared_weight) {
-  f::NetOp net;
+  ops::NetOp net;
   net.AddOp(f::OpRegistry::CreateOp("mul", {"X", "W"}, {"Out"}, {}));
   net.AddOp(f::OpRegistry::CreateOp("mul", {"Out", "W"}, {"FinalOut"}, {}));
   net.CompleteAddOp();
 
   auto bwd = f::Backward(net, {});
   ASSERT_TRUE(bwd->IsNetOp());
-  auto bwd_net = static_cast<f::NetOp *>(bwd.get());
+  auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
   ASSERT_EQ(3UL, bwd_net->ops_.size());
   ASSERT_EQ("add", bwd_net->ops_[2]->type_);
 }
@@ -285,7 +283,7 @@ TEST(Backward, op_all_input_are_not_need) {
   auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
   auto backward = f::Backward(*fwd, {"X", "b"});
   ASSERT_TRUE(backward->IsNetOp());
-  auto net = static_cast<f::NetOp *>(backward.get());
+  auto net = static_cast<ops::NetOp *>(backward.get());
   ASSERT_TRUE(net->ops_.empty());
 }
 
@@ -293,7 +291,7 @@ TEST(Backward, op_all_output_are_not_need) {
   auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
   auto backward = f::Backward(*fwd, {"Out"});
   ASSERT_TRUE(backward->IsNetOp());
-  auto net = static_cast<f::NetOp *>(backward.get());
+  auto net = static_cast<ops::NetOp *>(backward.get());
   ASSERT_TRUE(net->ops_.empty());
 }
 
@@ -301,7 +299,7 @@ TEST(Backward, op_part_of_output_are_not_need) {
   auto fwd = f::OpRegistry::CreateOp("many_output_op", {"X"}, {"Y", "Z"}, {});
   auto backward = f::Backward(*fwd, {"Z"});
   ASSERT_TRUE(backward->IsNetOp());
-  auto net = static_cast<f::NetOp *>(backward.get());
+  auto net = static_cast<ops::NetOp *>(backward.get());
   ASSERT_EQ(net->ops_.size(), 2UL);
 
   auto &fill_zero = *net->ops_[0];
@@ -309,17 +307,15 @@ TEST(Backward, op_part_of_output_are_not_need) {
   ASSERT_EQ(1UL, fill_zero.inputs_.size());
   ASSERT_EQ("Z", fill_zero.inputs_[0]);
   ASSERT_EQ(1UL, fill_zero.outputs_.size());
-  ASSERT_EQ("Z" + f::OperatorBase::ZERO_VAR_SUFFIX(), fill_zero.outputs_[0]);
+  ASSERT_EQ("Z" + f::kZeroVarSuffix, fill_zero.outputs_[0]);
 
   auto &d_many_out = *net->ops_[1];
   ASSERT_EQ("many_output_op_grad", d_many_out.type_);
   ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.inputs_.size());  // I/O/OG
-  ASSERT_EQ("Z" + f::OperatorBase::ZERO_VAR_SUFFIX(),
-            d_many_out.Input("z" + f::OperatorBase::GRAD_VAR_SUFFIX()));
-  ASSERT_EQ("Y" + f::OperatorBase::GRAD_VAR_SUFFIX(),
-            d_many_out.Input("y" + f::OperatorBase::GRAD_VAR_SUFFIX()));
-  ASSERT_EQ("X" + f::OperatorBase::GRAD_VAR_SUFFIX(),
-            d_many_out.Output("x" + f::OperatorBase::GRAD_VAR_SUFFIX()));
+  ASSERT_EQ("Z" + f::kZeroVarSuffix, d_many_out.Input("z" + f::kGradVarSuffix));
+  ASSERT_EQ("Y" + f::kGradVarSuffix, d_many_out.Input("y" + f::kGradVarSuffix));
+  ASSERT_EQ("X" + f::kGradVarSuffix,
+            d_many_out.Output("x" + f::kGradVarSuffix));
 }
 
 TEST(Backward, op_part_of_input_are_not_need) {
@@ -329,19 +325,17 @@ TEST(Backward, op_part_of_input_are_not_need) {
   ASSERT_EQ(grad_mul.type_, "mul_grad");
   ASSERT_EQ(grad_mul.inputs_.size(), 2UL + 1UL + 1UL);
   ASSERT_EQ(grad_mul.outputs_.size(), 2UL);
-  ASSERT_EQ(grad_mul.Output("A" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-            f::OperatorBase::EMPTY_VAR_NAME());
-  ASSERT_EQ(grad_mul.Output("B" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-            "b" + f::OperatorBase::GRAD_VAR_SUFFIX());
-  ASSERT_EQ(grad_mul.Input("Out" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-            "out" + f::OperatorBase::GRAD_VAR_SUFFIX());
+  ASSERT_EQ(grad_mul.Output("A" + f::kGradVarSuffix), f::kEmptyVarName);
+  ASSERT_EQ(grad_mul.Output("B" + f::kGradVarSuffix), "b" + f::kGradVarSuffix);
+  ASSERT_EQ(grad_mul.Input("Out" + f::kGradVarSuffix),
+            "out" + f::kGradVarSuffix);
   ASSERT_EQ(grad_mul.Input("A"), "a");
   ASSERT_EQ(grad_mul.Input("B"), "b");
   ASSERT_EQ(grad_mul.Input("Out"), "out");
 }
 
 TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
-  f::NetOp net;
+  ops::NetOp net;
   net.AddOp(f::OpRegistry::CreateOp("fc", {"x1", "w1", "b1"},
                                     {"mul_out1", "add_out1", "out1"}, {}));
   net.AddOp(f::OpRegistry::CreateOp("fc", {"out1", "w2", "b2"},
@@ -351,14 +345,13 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
   net.CompleteAddOp();
   auto backward = f::Backward(net, {"mul_out2", "tmp_out2", "out2"});
   ASSERT_TRUE(backward->IsNetOp());
-  auto bwd_net = static_cast<f::NetOp *>(backward.get());
+  auto bwd_net = static_cast<ops::NetOp *>(backward.get());
   ASSERT_EQ(bwd_net->ops_.size(), 3UL);
   auto &grad_fc = *bwd_net->ops_[0];
   EXPECT_EQ(grad_fc.inputs_.size(),
             3UL       /* external input number */
                 + 1UL /* external output number*/
                 + 1UL /* number of gradient of external output*/
-                - 1UL /*ignoreGradient varable number*/
                 + 2U /* internal variable number*/);
   EXPECT_EQ(grad_fc.outputs_.size(), 2UL       /* input number of mul*/
                                          + 2UL /* input number of rowwise_add */
@@ -367,23 +360,4 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
   EXPECT_EQ(bwd_net->ops_[1]->outputs_.size(), 0UL);
   EXPECT_EQ(bwd_net->ops_[2]->inputs_.size(), 0UL);
   EXPECT_EQ(bwd_net->ops_[2]->outputs_.size(), 0UL);
-
-  /*
-    EXPECT_EQ(grad_fc.Output("X" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-              f::OperatorBase::EMPTY_VAR_NAME());
-  EXPECT_EQ(grad_fc.Output("W" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-    "w3" + f::OperatorBase::GRAD_VAR_SUFFIX());
-  EXPECT_EQ(grad_fc.Output("b" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-    "b3" + f::OperatorBase::GRAD_VAR_SUFFIX());
-  EXPECT_EQ(grad_fc.Output("mul_result" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-  "mul_out3" + f::OperatorBase::GRAD_VAR_SUFFIX());
-
-  EXPECT_EQ(grad_fc.Input("Out" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-  "out3" + f::OperatorBase::GRAD_VAR_SUFFIX());
-  EXPECT_EQ(grad_fc.Input("X"), "out2");
-  EXPECT_EQ(grad_fc.Input("W"), "w3");
-  EXPECT_EQ(grad_fc.Input("mul_result"), "mul_out3");
-  EXPECT_EQ(grad_fc.Input("add_result"), "tmp_out3");
-  EXPECT_EQ(grad_fc.Input("Out"), "out3");
-  */
 }

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.cc

@@ -8,107 +8,95 @@ You may obtain a copy of the License at
 
 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. */
+WITHOpArgType::OUT WARRANTIES OR CONDITIONS OF ANY KOpArgType::IND, either
+express or implied. See the License for the specific language governing
+permissions and limitations under the License. */
 
 #include "paddle/framework/grad_op_builder.h"
+#include "paddle/framework/op_proto.pb.h"
 #include "paddle/framework/op_registry.h"
 
 namespace paddle {
 namespace framework {
 
-OperatorBase* GradOpBuilder::Build() {
-  BuildOpInOutArgList();
-  std::string grad_op_type = OpRegistry::grad_ops().at(op_.type_);
-  OperatorBase* grad_op = OpRegistry::op_creators().at(grad_op_type)();
-  grad_op->type_ = grad_op_type;
-  CompleteGradOp(grad_op);
-  return grad_op;
-}
+class OpRegistry;
+
+using VarIndexMap = std::unordered_map<std::string, int>;
 
-OpInOutArg* GradOpBuilder::BuildArg(const VarProto& var,
-                                    const VarIndexMap& var_map,
-                                    const std::vector<int>& format,
-                                    InOutType type) {
-  int idx = var_map.at(var.name());
-  int begin_idx = format.empty() ? idx : format.at(idx);
-  int end_idx = format.empty() ? idx + 1 : format.at(idx + 1);
-  return new OpInOutArg(var.name(), type, !var.ignore_gradient(), begin_idx,
-                        end_idx);
+enum class OpArgType { IN, OUT };
+
+static std::vector<int>* GetOpFormat(OperatorBase* op, const OpArgType& type) {
+  std::string key = type == OpArgType::IN ? "input_format" : "output_format";
+  return op->attrs_.count(key)
+             ? &boost::get<std::vector<int>>(op->attrs_.at(key))
+             : nullptr;
 }
 
-void GradOpBuilder::BuildOpInOutArgList() {
-  const OpProto& op_proto = OpRegistry::protos().at(op_.type_);
-  const auto& var_map = *(OpRegistry::VarIndexMaps().at(op_.type_));
-  const std::vector<int>& in_format =
-      op_.attrs_.count("input_format")
-          ? op_.GetAttr<std::vector<int>>("input_format")
-          : std::vector<int>();
-  const std::vector<int>& out_format =
-      op_.attrs_.count("output_format")
-          ? op_.GetAttr<std::vector<int>>("output_format")
-          : std::vector<int>();
-  for (const auto& var : op_proto.inputs()) {
-    arg_list_.emplace_back(
-        std::shared_ptr<OpInOutArg>(BuildArg(var, var_map, in_format, IN)));
-  }
-  for (const auto& var : op_proto.outputs()) {
-    arg_list_.emplace_back(
-        std::shared_ptr<OpInOutArg>(BuildArg(var, var_map, out_format, OUT)));
-  }
+static const std::vector<int>* GetOpFormat(const OperatorBase* op,
+                                           const OpArgType& type) {
+  std::string key = type == OpArgType::IN ? "input_format" : "output_format";
+  return op->attrs_.count(key)
+             ? &boost::get<std::vector<int>>(op->attrs_.at(key))
+             : nullptr;
 }
 
-void GradOpBuilder::AddArgIntoGradOp(const OpInOutArg* arg,
-                                     std::vector<std::string>& in_out,
-                                     std::vector<int>& format,
-                                     VarIndexMap* varmap, int& idx,
-                                     bool is_grad) const {
-  std::string var_name = arg->proto_name_;
-  if (is_grad) {
-    var_name += OperatorBase::GRAD_VAR_SUFFIX();
-  }
-  (*varmap)[var_name] = idx++;
-  size_t pre_sz = in_out.size();
-  auto base_it = arg->type_ == IN ? op_.inputs_.begin() : op_.outputs_.begin();
-  std::copy(base_it + arg->begin_idx_, base_it + arg->end_idx_,
-            std::back_inserter(in_out));
-  if (is_grad) {
-    for (size_t i = pre_sz; i < in_out.size(); ++i) {
-      in_out[i] += OperatorBase::GRAD_VAR_SUFFIX();
+static void TransOpArg(const OperatorBase* src_op, OperatorBase* dst_op,
+                       const OpArgType& src_type, const OpArgType& dst_type,
+                       int& idx, bool is_grad) {
+  const std::vector<std::string>& src_inout =
+      src_type == OpArgType::IN ? src_op->inputs_ : src_op->outputs_;
+  const std::vector<int>* src_format = GetOpFormat(src_op, src_type);
+
+  std::vector<std::string>& dst_inout =
+      dst_type == OpArgType::IN ? dst_op->inputs_ : dst_op->outputs_;
+  std::vector<int>* dst_format = GetOpFormat(dst_op, dst_type);
+  const OpProto& proto = OpRegistry::protos().at(src_op->type_);
+  const auto& src_arg_list =
+      src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
+
+  for (const auto& arg : src_arg_list) {
+    std::string src_name = arg.name();
+    std::string dst_name = is_grad ? src_name + kGradVarSuffix : src_name;
+    (*dst_op->in_out_idxs_)[dst_name] = idx++;
+    int src_arg_idx = src_op->in_out_idxs_->at(src_name);
+    int src_begin =
+        src_format == nullptr ? src_arg_idx : src_format->at(src_arg_idx);
+    int src_end = src_format == nullptr ? src_arg_idx + 1
+                                        : src_format->at(src_arg_idx + 1);
+    for (int i = src_begin; i < src_end; ++i) {
+      std::string s =
+          is_grad ? src_inout[i] + kGradVarSuffix
+                  : (arg.ignore_gradient() ? kEmptyVarName : src_inout[i]);
+      dst_inout.emplace_back(s);
+    }
+    if (dst_format != nullptr) {
+      dst_format->push_back(dst_inout.size());
     }
   }
-  format.push_back(in_out.size());
 }
 
-void GradOpBuilder::CompleteGradOp(OperatorBase* grad_op) const {
-  grad_op->attrs_ = op_.attrs_;
+OperatorBase* BuildGradOp(const OperatorBase* op) {
+  std::string grad_op_type = OpRegistry::grad_ops().at(op->type_);
+  OperatorBase* grad_op = OpRegistry::op_creators().at(grad_op_type)();
+  grad_op->type_ = grad_op_type;
+  grad_op->attrs_ = op->attrs_;
   grad_op->attrs_.erase("input_format");
   grad_op->attrs_.erase("output_format");
-  VarIndexMap* grad_varmap = new VarIndexMap();
+  if (GetOpFormat(op, OpArgType::IN) != nullptr) {
+    grad_op->attrs_["output_format"] = std::vector<int>({0});
+  }
+  if (GetOpFormat(op, OpArgType::IN) != nullptr ||
+      GetOpFormat(op, OpArgType::OUT) != nullptr) {
+    grad_op->attrs_["input_format"] = std::vector<int>({0});
+  }
+  grad_op->in_out_idxs_.reset(new VarIndexMap());
   int in_idx = 0;
   int out_idx = 0;
-  std::vector<int> in_format({0});
-  std::vector<int> out_format({0});
-  for (const auto& arg : arg_list_) {
-    // op_'s inputs_ and outputs_
-    if (arg->needed_in_grad_) {
-      AddArgIntoGradOp(arg.get(), grad_op->inputs_, in_format, grad_varmap,
-                       in_idx, false);
-    }
-    if (arg->type_ == IN) {
-      // gradients of op_'s inputs_
-      AddArgIntoGradOp(arg.get(), grad_op->outputs_, out_format, grad_varmap,
-                       out_idx, true);
-    } else {
-      // gradients of op_'s outputs_
-      AddArgIntoGradOp(arg.get(), grad_op->inputs_, in_format, grad_varmap,
-                       in_idx, true);
-    }
-  }
-  grad_op->attrs_["input_format"] = in_format;
-  grad_op->attrs_["output_format"] = out_format;
-  grad_op->in_out_idxs_.reset(grad_varmap);
+  TransOpArg(op, grad_op, OpArgType::IN, OpArgType::IN, in_idx, false);   // I
+  TransOpArg(op, grad_op, OpArgType::OUT, OpArgType::IN, in_idx, false);  // G
+  TransOpArg(op, grad_op, OpArgType::OUT, OpArgType::IN, in_idx, true);   // OG
+  TransOpArg(op, grad_op, OpArgType::IN, OpArgType::OUT, out_idx, true);  // IG
+  return grad_op;
 }
 
 }  // namespace framework

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/op_proto.pb.h"
 #include "paddle/framework/operator.h"
 
 namespace paddle {
 namespace framework {
-class OpRegistry;
-
-enum InOutType { IN, OUT };
-
-struct OpInOutArg {
-  OpInOutArg(const std::string& proto_name, const InOutType& type,
-             bool needed_in_grad, size_t begin_idx, size_t end_idx)
-      : proto_name_(proto_name),
-        type_(type),
-        needed_in_grad_(needed_in_grad),
-        begin_idx_(begin_idx),
-        end_idx_(end_idx) {}
-
-  std::string proto_name_;
-  InOutType type_;
-  bool needed_in_grad_;
-  size_t begin_idx_;
-  size_t end_idx_;
-};
-
-class GradOpBuilder {
-  using VarIndexMap = std::unordered_map<std::string, int>;
-
- public:
-  GradOpBuilder(const OperatorBase& op) : op_(op) {}
-  OperatorBase* Build();
-
- private:
-  OpInOutArg* BuildArg(const VarProto& var, const VarIndexMap& var_map,
-                       const std::vector<int>& format, InOutType type);
-  void BuildOpInOutArgList();
-  void AddArgIntoGradOp(const OpInOutArg* arg, std::vector<std::string>& in_out,
-                        std::vector<int>& format, VarIndexMap* varmap, int& idx,
-                        bool is_grad) const;
-  void CompleteGradOp(OperatorBase* grad_op) const;
-  const OperatorBase& op_;
-  std::vector<std::shared_ptr<OpInOutArg>> arg_list_;
-};
+
+OperatorBase* BuildGradOp(const OperatorBase* op);
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/grad_op_builder_test.cc

@@ -8,10 +8,49 @@ USE_OP(add_two);
 namespace paddle {
 namespace framework {
 
+class NOP : public OperatorBase {
+ public:
+  void InferShape(const Scope &scope) const override {}
+  void Run(const Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {}
+};
+
+class MutiInOutOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  MutiInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("In1", "a single input");
+    AddInput("In2_mult", "a multiple input").SetMultiple();
+    AddInput("In3", "another single input");
+    AddOutput("Out1", "a single output");
+    AddOutput("Out2_mult", "a multiple output").SetMultiple();
+    AddComment("test op with multiple inputs and outputs");
+  }
+};
+
+class IOIgnoredOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  IOIgnoredOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("In1", "a single input");
+    AddInput("In2_mult", "a multiple input").SetMultiple().IgnoreGradient();
+    AddInput("In3_mult", "another multiple input").SetMultiple();
+    AddOutput("Out1_mult", "a multiple output").SetMultiple();
+    AddOutput("Out2", "a single output").IgnoreGradient();
+    AddComment("op with inputs and outputs ignored in gradient calculating");
+  }
+};
+
+}  // namespace framework
+}  // namespace paddle
+
+namespace f = paddle::framework;
+
 TEST(GradOpBuilder, AddTwo) {
-  std::shared_ptr<OperatorBase> add_op(
-      OpRegistry::CreateOp("add_two", {"x", "y"}, {"out"}, {}));
-  std::shared_ptr<OperatorBase> grad_add_op = OpRegistry::CreateGradOp(*add_op);
+  std::shared_ptr<f::OperatorBase> add_op(
+      f::OpRegistry::CreateOp("add_two", {"x", "y"}, {"out"}, {}));
+  std::shared_ptr<f::OperatorBase> grad_add_op =
+      f::OpRegistry::CreateGradOp(*add_op);
   EXPECT_EQ(static_cast<int>(grad_add_op->inputs_.size()), 4);
   EXPECT_EQ(static_cast<int>(grad_add_op->outputs_.size()), 2);
   EXPECT_EQ(grad_add_op->Input("X"), "x");
@@ -22,5 +61,77 @@ TEST(GradOpBuilder, AddTwo) {
   EXPECT_EQ(grad_add_op->Output("Y@GRAD"), "y@GRAD");
 }
 
-}  // namespace framework
-}  // namespace paddle
+REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker);
+REGISTER_GRADIENT_OP(mult_io, mult_io_grad, f::NOP);
\ No newline at end of file
+REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker);
+REGISTER_GRADIENT_OP(io_ignored, io_ignored_grad, f::NOP);
+
+TEST(GradOpBuilder, MutiInOut) {
+  f::AttributeMap attrs{{"input_format", std::vector<int>{0, 1, 4, 5}},
+                        {"output_format", std::vector<int>{0, 1, 3}}};
+  std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
+      "mult_io", {"in1", "in2_1", "in2_2", "in2_3", "in3"},
+      {"out1", "out2_1", "out2_2"}, attrs));
+  std::shared_ptr<f::OperatorBase> grad_test_op =
+      f::OpRegistry::CreateGradOp(*test_op);
+
+  ASSERT_EQ(grad_test_op->inputs_.size(), 5UL + 3UL + 3UL);
+  EXPECT_EQ(grad_test_op->Input("In1"), "in1");
+  EXPECT_EQ(grad_test_op->Inputs("In2_mult"),
+            std::vector<std::string>({"in2_1", "in2_2", "in2_3"}));
+  EXPECT_EQ(grad_test_op->Input("In3"), "in3");
+  EXPECT_EQ(grad_test_op->Input("Out1"), "out1");
+  EXPECT_EQ(grad_test_op->Inputs("Out2_mult"),
+            std::vector<std::string>({"out2_1", "out2_2"}));
+  EXPECT_EQ(grad_test_op->Input("Out1" + f::kGradVarSuffix),
+            "out1" + f::kGradVarSuffix);
+  EXPECT_EQ(grad_test_op->Inputs("Out2_mult" + f::kGradVarSuffix),
+            std::vector<std::string>(
+                {"out2_1" + f::kGradVarSuffix, "out2_2" + f::kGradVarSuffix}));
+
+  ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
+  EXPECT_EQ(grad_test_op->Output("In1" + f::kGradVarSuffix),
+            "in1" + f::kGradVarSuffix);
+  EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix),
+            std::vector<std::string>({"in2_1" + f::kGradVarSuffix,
+                                      "in2_2" + f::kGradVarSuffix,
+                                      "in2_3" + f::kGradVarSuffix}));
+  EXPECT_EQ(grad_test_op->Output("In3" + f::kGradVarSuffix),
+            "in3" + f::kGradVarSuffix);
+}
+
+TEST(GradOpBuilder, IOIgnoredInGradient) {
+  f::AttributeMap attrs{{"input_format", std::vector<int>{0, 1, 3, 5}},
+                        {"output_format", std::vector<int>{0, 2, 3}}};
+  std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
+      "io_ignored", {"in1", "in2_1", "in2_2", "in3_1", "in3_2"},
+      {"out1_1", "out1_2", "out2"}, attrs));
+  std::shared_ptr<f::OperatorBase> grad_test_op =
+      f::OpRegistry::CreateGradOp(*test_op);
+
+  // 'In2' and 'Out2' are ignored in gradient calculating
+  ASSERT_EQ(grad_test_op->inputs_.size(), 5UL + 3UL + 3UL);
+  EXPECT_EQ(grad_test_op->Input("In1"), "in1");
+  EXPECT_EQ(grad_test_op->Inputs("In2_mult"),
+            std::vector<std::string>({f::kEmptyVarName, f::kEmptyVarName}));
+  EXPECT_EQ(grad_test_op->Inputs("In3_mult"),
+            std::vector<std::string>({"in3_1", "in3_2"}));
+  EXPECT_EQ(grad_test_op->Inputs("Out1_mult"),
+            std::vector<std::string>({"out1_1", "out1_2"}));
+  EXPECT_EQ(grad_test_op->Input("Out2"), f::kEmptyVarName);
+  EXPECT_EQ(grad_test_op->Inputs("Out1_mult" + f::kGradVarSuffix),
+            std::vector<std::string>(
+                {"out1_1" + f::kGradVarSuffix, "out1_2" + f::kGradVarSuffix}));
+  EXPECT_EQ(grad_test_op->Input("Out2" + f::kGradVarSuffix),
+            "out2" + f::kGradVarSuffix);
+
+  ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
+  EXPECT_EQ(grad_test_op->Output("In1" + f::kGradVarSuffix),
+            "in1" + f::kGradVarSuffix);
+  EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix),
+            std::vector<std::string>(
+                {"in2_1" + f::kGradVarSuffix, "in2_2" + f::kGradVarSuffix}));
+  EXPECT_EQ(grad_test_op->Outputs("In3_mult" + f::kGradVarSuffix),
+            std::vector<std::string>(
+                {"in3_1" + f::kGradVarSuffix, "in3_2" + f::kGradVarSuffix}));
+}

paddle/framework/op_desc.proto

@@ -15,7 +15,7 @@ limitations under the License. */
 syntax="proto2";
 package paddle.framework;
 
-import "attr_type.proto";
+import "attribute.proto";
 
 // AttrDesc is used to describe Attributes of an Operator. It contain's
 // name, type, and value of Attribute.

paddle/framework/op_proto.proto

@@ -21,7 +21,7 @@ limitations under the License. */
 syntax="proto2";
 package paddle.framework;
 
-import "attr_type.proto";
+import "attribute.proto";
 
 // Attribute protocol message for 3rd-party language binding.
 // It will store the Op support what attribute and what type.

paddle/framework/op_registry.cc

@@ -14,37 +14,8 @@ limitations under the License. */
 
 #include <paddle/framework/op_registry.h>
 
-namespace paddle {
-namespace framework {
-
-template <>
-void AttrTypeHelper::SetAttrType<int>(AttrProto* attr) {
-  attr->set_type(paddle::framework::AttrType::INT);
-}
-
-template <>
-void AttrTypeHelper::SetAttrType<float>(AttrProto* attr) {
-  attr->set_type(paddle::framework::AttrType::FLOAT);
-}
-
-template <>
-void AttrTypeHelper::SetAttrType<std::string>(AttrProto* attr) {
-  attr->set_type(paddle::framework::AttrType::STRING);
-}
+#include <vector>
 
-template <>
-void AttrTypeHelper::SetAttrType<std::vector<int>>(AttrProto* attr) {
-  attr->set_type(paddle::framework::AttrType::INTS);
-}
-
-template <>
-void AttrTypeHelper::SetAttrType<std::vector<float>>(AttrProto* attr) {
-  attr->set_type(paddle::framework::AttrType::FLOATS);
-}
-
-template <>
-void AttrTypeHelper::SetAttrType<std::vector<std::string>>(AttrProto* attr) {
-  attr->set_type(paddle::framework::AttrType::STRINGS);
-}
-}  // namespace framework
+namespace paddle {
+namespace framework {}  // namespace framework
 }  // namespace paddle

paddle/framework/op_registry.h

@@ -19,7 +19,7 @@ limitations under the License. */
 #include <type_traits>
 #include <unordered_map>
 #include <unordered_set>
-#include "paddle/framework/attr_checker.h"
+#include "paddle/framework/attribute.h"
 #include "paddle/framework/grad_op_builder.h"
 #include "paddle/framework/op_desc.pb.h"
 #include "paddle/framework/scope.h"
@@ -27,49 +27,6 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-// helper class to set attribute type
-struct AttrTypeHelper {
-  template <typename T>
-  static void SetAttrType(AttrProto* attr);
-
-  static Attribute GetAttrValue(const AttrDesc& attr_desc) {
-    switch (attr_desc.type()) {
-      case paddle::framework::AttrType::INT: {
-        return attr_desc.i();
-      }
-      case paddle::framework::AttrType::FLOAT: {
-        return attr_desc.f();
-      }
-      case paddle::framework::AttrType::STRING: {
-        return attr_desc.s();
-      }
-      case paddle::framework::AttrType::INTS: {
-        std::vector<int> val(attr_desc.ints_size());
-        for (int i = 0; i < attr_desc.ints_size(); ++i) {
-          val[i] = attr_desc.ints(i);
-        }
-        return val;
-      }
-      case paddle::framework::AttrType::FLOATS: {
-        std::vector<float> val(attr_desc.floats_size());
-        for (int i = 0; i < attr_desc.floats_size(); ++i) {
-          val[i] = attr_desc.floats(i);
-        }
-        return val;
-      }
-      case paddle::framework::AttrType::STRINGS: {
-        std::vector<std::string> val(attr_desc.strings_size());
-        for (int i = 0; i < attr_desc.strings_size(); ++i) {
-          val[i] = attr_desc.strings(i);
-        }
-        return val;
-      }
-    }
-    PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !");
-    return boost::blank();
-  }
-};
-
 // this class not only make proto but also init attribute checkers.
 class OpProtoAndCheckerMaker {
  public:
@@ -136,7 +93,7 @@ class OpProtoAndCheckerMaker {
     *attr->mutable_name() = name;
     *attr->mutable_comment() = comment;
     attr->set_generated(generated);
-    AttrTypeHelper::SetAttrType<T>(attr);
+    attr->set_type(AttrTypeID<T>());
     return op_checker_->AddAttrChecker<T>(name);
   }
 
@@ -297,7 +254,7 @@ class OpRegistry {
 
     AttributeMap attrs;
     for (auto& attr : op_desc.attrs()) {
-      attrs[attr.name()] = AttrTypeHelper::GetAttrValue(attr);
+      attrs[attr.name()] = GetAttrValue(attr);
     }
 
     return CreateOp(op_desc.type(), inputs, outputs, attrs);
@@ -306,8 +263,7 @@ class OpRegistry {
   static std::shared_ptr<OperatorBase> CreateGradOp(const OperatorBase& op) {
     PADDLE_ENFORCE(!op.IsNetOp(),
                    "Use framework::Backward to get backward ops");
-    GradOpBuilder builder(op);
-    std::shared_ptr<OperatorBase> grad_op(builder.Build());
+    std::shared_ptr<OperatorBase> grad_op(BuildGradOp(&op));
     grad_op->Init();
     return grad_op;
   }
@@ -315,7 +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_;
@@ -337,12 +293,12 @@ 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);
     for (auto& outname : op->outputs_) {
-      if (outname == OperatorBase::TMP_VAR_NAME()) {
+      if (outname == kTempVarName) {
         outname += op->type_;
         outname += "@";
         outname += std::to_string(gUniqId.fetch_add(1));
@@ -354,7 +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

@@ -20,7 +20,7 @@ limitations under the License. */
 #include <unordered_map>
 #include <vector>
 
-#include "paddle/framework/attr_checker.h"
+#include "paddle/framework/attribute.h"
 #include "paddle/framework/op_desc.pb.h"
 #include "paddle/framework/op_proto.pb.h"
 #include "paddle/framework/scope.h"
@@ -32,9 +32,29 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+/// If a variable is a empty variable, that name will be used.
+const std::string kEmptyVarName = "@EMPTY@";
+
+/// If a variable is a temporary variable, that name will be set in Python,
+/// but it will be convert to a unique name in scope after OpCreator.
+const std::string kTempVarName = "@TEMP@";
+
+/// If a variable's name has a certain suffix, it means that the
+/// variable is the gradient of another varibale.
+/// e.g. Variable "x@GRAD" is the gradient of varibale "x".
+const std::string kGradVarSuffix = "@GRAD";
+
+/// Variables with this suffix are supposed to be filled up with zeros.
+const std::string kZeroVarSuffix = "@ZERO";
+
+inline std::string GradVarName(const std::string& var_name) {
+  return var_name + kGradVarSuffix;
+}
+
 class OperatorBase;
 class InferShapeContext;
 class ExecutionContext;
+
 /**
  * OperatorBase has the basic element that Net will call to do computation.
  * Only CreateOperator from OpRegistry will new Operator directly. User
@@ -43,21 +63,6 @@ class ExecutionContext;
  */
 class OperatorBase {
  public:
-  /// If a variable is a empty variable, that name will be used.
-  static std::string EMPTY_VAR_NAME() { return "@EMPTY@"; }
-
-  /// If a variable is a temporary variable, that name will be set in Python,
-  /// but it will be convert to a unique name in scope after OpCreator.
-  static std::string TMP_VAR_NAME() { return "@TEMP@"; }
-
-  /// If a variable's name has a certain suffix, it means that the
-  /// variable is the gradient of another varibale.
-  /// e.g. Variable "x@GRAD" is the gradient of varibale "x".
-  static std::string GRAD_VAR_SUFFIX() { return "@GRAD"; }
-
-  /// Variables with this suffix are supposed to be filled up with zeros.
-  static std::string ZERO_VAR_SUFFIX() { return "@ZERO"; }
-
   virtual ~OperatorBase() {}
 
   template <typename T>
@@ -280,7 +285,7 @@ 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/pybind/pybind.cc → paddle/framework/pybind.cc

@@ -4,7 +4,7 @@ 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
+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,
@@ -17,19 +17,19 @@ limitations under the License. */
 #include <vector>
 
 #include "paddle/framework/backward.h"
-#include "paddle/framework/net.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/scope.h"
+#include "paddle/framework/tensor_py.h"
+#include "paddle/operators/net_op.h"
+#include "paddle/operators/type_alias.h"
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/place.h"
-#include "paddle/pybind/tensor_bind.h"
 #include "pybind11/numpy.h"
 #include "pybind11/pybind11.h"
 #include "pybind11/stl.h"
 
 namespace py = pybind11;
-namespace pd = paddle::framework;
 
 USE_OP(add_two);
 USE_OP(onehot_cross_entropy);
@@ -41,17 +41,18 @@ USE_OP(sigmoid);
 USE_OP(softmax);
 USE_OP(rowwise_add);
 USE_OP_WITHOUT_KERNEL(recurrent_op);
-
+namespace paddle {
+namespace framework {
 template <typename ClassType>
-void ExposeOperator(ClassType& m) {
+void ExposeOperator(ClassType &m) {
   m.def("infer_shape", &ClassType::type::InferShape)
       .def("run", &ClassType::type::Run)
       .def("type",
-           [](const typename ClassType::type& op) -> std::string {
+           [](const typename ClassType::type &op) -> std::string {
              return op.type_;
            })
       .def("outputs",
-           [](const typename ClassType::type& op) -> std::vector<std::string> {
+           [](const typename ClassType::type &op) -> std::vector<std::string> {
              return op.outputs_;
            })
       .def("__str__", &ClassType::type::DebugString);
@@ -73,80 +74,81 @@ bool IsCompileGPU() {
 PYBIND11_PLUGIN(core) {
   py::module m("core", "C++ core of PaddlePaddle");
 
-  py::class_<pd::Tensor>(m, "Tensor", py::buffer_protocol())
-      .def_buffer([](pd::Tensor& self) -> py::buffer_info {
-        return paddle::pybind::CastToPyBuffer(self);
-      })
+  py::class_<Tensor>(m, "Tensor", py::buffer_protocol())
+      .def_buffer(
+          [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
       .def("get_dims",
-           [](const pd::Tensor& self) { return pd::vectorize(self.dims()); })
+           [](const Tensor &self) { return vectorize(self.dims()); })
       .def("set_dims",
-           [](pd::Tensor& self, const std::vector<int>& dim) {
-             self.Resize(pd::make_ddim(dim));
+           [](Tensor &self, const std::vector<int> &dim) {
+             self.Resize(make_ddim(dim));
            })
       .def("alloc_float",
-           [](pd::Tensor& self, paddle::platform::GPUPlace& place) {
+           [](Tensor &self, paddle::platform::GPUPlace &place) {
              self.mutable_data<float>(place);
            })
       .def("alloc_float",
-           [](pd::Tensor& self, paddle::platform::CPUPlace& place) {
+           [](Tensor &self, paddle::platform::CPUPlace &place) {
              self.mutable_data<float>(place);
            })
       .def("alloc_int",
-           [](pd::Tensor& self, paddle::platform::CPUPlace& place) {
+           [](Tensor &self, paddle::platform::CPUPlace &place) {
              self.mutable_data<int>(place);
            })
       .def("alloc_int",
-           [](pd::Tensor& self, paddle::platform::GPUPlace& place) {
+           [](Tensor &self, paddle::platform::GPUPlace &place) {
              self.mutable_data<int>(place);
            })
-      .def("set", paddle::pybind::PyCPUTensorSetFromArray<float>)
-      .def("set", paddle::pybind::PyCPUTensorSetFromArray<int>)
+      .def("set", PyCPUTensorSetFromArray<float>)
+      .def("set", PyCPUTensorSetFromArray<int>)
 #ifndef PADDLE_ONLY_CPU
-      .def("set", paddle::pybind::PyCUDATensorSetFromArray<float>)
-      .def("set", paddle::pybind::PyCUDATensorSetFromArray<int>)
+      .def("set", PyCUDATensorSetFromArray<float>)
+      .def("set", PyCUDATensorSetFromArray<int>)
 #endif
-      .def("shape",
-           [](pd::Tensor& self) { return pd::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_<pd::Variable>(m, "Variable", R"DOC(Variable Class.
+  py::class_<Variable>(m, "Variable", R"DOC(Variable Class.
 
 All parameter, weight, gradient are variables in Paddle.
 )DOC")
-      .def("is_int", [](const pd::Variable& var) { return var.IsType<int>(); })
+      .def("is_int", [](const Variable &var) { return var.IsType<int>(); })
       .def("set_int",
-           [](pd::Variable& var, int val) -> void {
-             *var.GetMutable<int>() = val;
-           })
-      .def("get_int",
-           [](const pd::Variable& var) -> int { return var.Get<int>(); })
+           [](Variable &var, int val) -> void { *var.GetMutable<int>() = val; })
+      .def("get_int", [](const Variable &var) -> int { return var.Get<int>(); })
       .def("get_tensor",
-           [](pd::Variable& self) -> pd::Tensor* {
-             return self.GetMutable<pd::Tensor>();
-           },
+           [](Variable &self) -> Tensor * { return self.GetMutable<Tensor>(); },
            py::return_value_policy::reference)
       .def("get_net",
-           [](pd::Variable& self) -> pd::NetOp* {
-             return self.GetMutable<pd::NetOp>();
+           [](Variable &self) -> ops::NetOp * {
+             return self.GetMutable<ops::NetOp>();
            },
            py::return_value_policy::reference);
 
-  py::class_<pd::Scope>(m, "Scope", "")
+  py::class_<Scope>(m, "Scope", "")
       .def("new_var",
-           [](pd::Scope& self, const std::string& name) -> pd::Variable* {
+           [](Scope &self, const std::string &name) -> Variable * {
              return self.NewVar(name);
            },
            py::return_value_policy::reference)
-      .def("find_var", &pd::Scope::FindVar, py::return_value_policy::reference)
+      .def("find_var", &Scope::FindVar, py::return_value_policy::reference)
       .def(py::init<>())
-      .def("new_scope",
-           [](pd::Scope& self) -> pd::Scope* { return &self.NewScope(); },
+      .def("new_scope", [](Scope &self) -> Scope * { return &self.NewScope(); },
            py::return_value_policy::reference)
-      .def("drop_kids", &pd::Scope::DropKids);
+      .def("drop_kids", &Scope::DropKids);
 
   //! @note: Be careful! PyBind will return std::string as an unicode, not
   //! Python str. If you want a str object, you should cast them in Python.
   m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
-    auto& protos = pd::OpRegistry::protos();
+    auto &protos = OpRegistry::protos();
     std::vector<py::bytes> ret_values;
     for (auto it = protos.begin(); it != protos.end(); ++it) {
       PADDLE_ENFORCE(it->second.IsInitialized(),
@@ -161,8 +163,8 @@ All parameter, weight, gradient are variables in Paddle.
   m.def_submodule(
        "var_names",
        "The module will return special predefined variable name in Paddle")
-      .def("empty", pd::OperatorBase::EMPTY_VAR_NAME)
-      .def("temp", pd::OperatorBase::TMP_VAR_NAME);
+      .def("empty", []() { return kEmptyVarName; })
+      .def("temp", []() { return kTempVarName; });
   // clang-format off
   py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
       .def_static("create",
@@ -185,43 +187,45 @@ All parameter, weight, gradient are variables in Paddle.
 
   py::class_<paddle::platform::CPUPlace>(m, "CPUPlace").def(py::init<>());
 
-  py::class_<pd::OperatorBase, std::shared_ptr<pd::OperatorBase>> operator_base(
+  py::class_<OperatorBase, std::shared_ptr<OperatorBase>> operator_base(
       m, "Operator");
 
   operator_base.def_static("create", [](py::bytes protobin) {
-    pd::OpDesc desc;
+    OpDesc desc;
     PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
                    "Cannot parse user input to OpDesc");
     PADDLE_ENFORCE(desc.IsInitialized(),
                    "User OpDesc is not initialized, reason %s",
                    desc.InitializationErrorString());
-    return pd::OpRegistry::CreateOp(desc);
+    return OpRegistry::CreateOp(desc);
   });
 
   operator_base.def("backward",
-                    [](const pd::OperatorBase& forwardOp,
-                       const std::unordered_set<std::string>& no_grad_vars) {
-                      return pd::Backward(forwardOp, no_grad_vars);
+                    [](const OperatorBase &forwardOp,
+                       const std::unordered_set<std::string> &no_grad_vars) {
+                      return Backward(forwardOp, no_grad_vars);
                     });
 
   ExposeOperator(operator_base);
 
-  py::class_<pd::NetOp, std::shared_ptr<pd::NetOp>> net(m, "Net");
+  py::class_<ops::NetOp, std::shared_ptr<ops::NetOp>> net(m, "Net");
 
   net.def_static("create",
-                 []() -> std::shared_ptr<pd::NetOp> {
-                   auto retv = std::make_shared<pd::NetOp>();
+                 []() -> std::shared_ptr<ops::NetOp> {
+                   auto retv = std::make_shared<ops::NetOp>();
                    retv->type_ = "plain_net";
                    return retv;
                  })
-      .def("add_op", &pd::NetOp::AddOp)
-      .def("add_op",
-           [](pd::NetOp& self, const std::shared_ptr<pd::NetOp>& net) -> void {
-             self.AddOp(std::static_pointer_cast<pd::OperatorBase>(net));
-           })
-      .def("complete_add_op", &pd::NetOp::CompleteAddOp)
+      .def("add_op", &ops::NetOp::AddOp)
+      .def(
+          "add_op",
+          [](ops::NetOp &self, const std::shared_ptr<ops::NetOp> &net) -> void {
+            self.AddOp(std::static_pointer_cast<OperatorBase>(net));
+          })
+      .def("complete_add_op", &ops::NetOp::CompleteAddOp)
       .def("complete_add_op",
-           [](std::shared_ptr<pd::NetOp>& self) { self->CompleteAddOp(); });
+           [](std::shared_ptr<ops::NetOp> &self) { self->CompleteAddOp(); });
+
   ExposeOperator(net);
 
   m.def("unique_integer", UniqueIntegerGenerator);
@@ -230,3 +234,5 @@ All parameter, weight, gradient are variables in Paddle.
 
   return m.ptr();
 }
+}  // namespace framework
+}  // namespace paddle

paddle/framework/tensor.h

@@ -26,19 +26,17 @@ limitations under the License. */
 #include "unsupported/Eigen/CXX11/Tensor"
 
 namespace paddle {
-namespace pybind {
-namespace details {  // forward declare
-template <bool less, size_t i, typename... args>
-struct CastToPyBufferImpl;
-}  // namespace details
-}  // namespace pybind
 
 namespace framework {
+namespace details {
+template <bool less, size_t i, typename... args>
+struct CastToPyBufferImpl;
+}
 
 class Tensor {
  public:
   template <bool less, size_t i, typename... args>
-  friend struct paddle::pybind::details::CastToPyBufferImpl;
+  friend struct details::CastToPyBufferImpl;
 
   template <typename T, size_t D, int MajorType, typename IndexType>
   friend struct EigenTensor;

paddle/pybind/tensor_bind.h → paddle/framework/tensor_py.h

@@ -23,7 +23,7 @@ namespace py = pybind11;
 
 namespace paddle {
 
-namespace pybind {
+namespace framework {
 
 namespace details {
 
@@ -63,11 +63,8 @@ struct CastToPyBufferImpl<true, I, ARGS...> {
       }
       return py::buffer_info(
           dst_tensor.mutable_data<CUR_TYPE>(dst_tensor.holder_->place()),
-          sizeof(CUR_TYPE),
-          py::format_descriptor<CUR_TYPE>::format(),
-          (size_t)framework::arity(dst_tensor.dims()),
-          dims_outside,
-          strides);
+          sizeof(CUR_TYPE), py::format_descriptor<CUR_TYPE>::format(),
+          (size_t)framework::arity(dst_tensor.dims()), dims_outside, strides);
     } else {
       constexpr bool less = I + 1 < std::tuple_size<std::tuple<ARGS...>>::value;
       return CastToPyBufferImpl<less, I + 1, ARGS...>()(tensor);
@@ -110,8 +107,8 @@ void PyCUDATensorSetFromArray(
 
   self.Resize(framework::make_ddim(dims));
   auto *dst = self.mutable_data<T>(place);
-  paddle::platform::GpuMemcpySync(
-      dst, array.data(), sizeof(T) * array.size(), cudaMemcpyHostToDevice);
+  paddle::platform::GpuMemcpySync(dst, array.data(), sizeof(T) * array.size(),
+                                  cudaMemcpyHostToDevice);
 }
 #endif
 

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/CMakeLists.txt

 # gserver pacakge unittests
 
+file(GLOB_RECURSE GSERVER_HEADER RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.h")
+file(GLOB_RECURSE GSERVER_SOURCES RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.cpp")
+add_style_check_target(paddle_gserver ${GSERVER_SOURCES})
+add_style_check_target(paddle_gserver ${GSERVER_HEADER})
+
 ################### test_ProtoDataProvider ############
 add_unittest_without_exec(test_ProtoDataProvider
     test_ProtoDataProvider.cpp)
@@ -50,7 +55,7 @@ add_unittest_without_exec(test_DetectionOutput
     test_DetectionOutput.cpp
     LayerGradUtil.cpp)
 
-add_test(NAME test_DetectionOutput 
+add_test(NAME test_DetectionOutput
     COMMAND test_DetectionOutput)
 ################# test_ConvUnify #######################
 add_unittest_without_exec(test_ConvUnify

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/math/MathUtils.cpp

@@ -25,7 +25,7 @@ namespace paddle {
  */
 void sparseRand(
     int* major, int* minor, int nnz, int majorLen, int minorMax, bool useGpu) {
-  CHECK(size_t(nnz) > size_t(1));
+  CHECK(size_t(nnz) >= size_t(1));
   int* cpuMajor;
   int* cpuMinor;
   CpuIVector cpuMinorVec(nnz);

paddle/math/tests/test_matrixCompare.cpp

@@ -79,8 +79,8 @@ void testMatrixMaxSequence(int batchSize, int inputDim) {
 }
 
 TEST(Matrix, maxSequence) {
-  for (auto batchSize : {1, 10, 128, 1000, 6000}) {
-    for (auto inputDim : {1, 32, 100, 512}) {
+  for (auto batchSize : {1, 3, 997}) {   // prime numbers close to 1, 4, 1024
+    for (auto inputDim : {1, 7, 131}) {  // prime numbers close to 1, 8, 128
       VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim;
       testMatrixMaxSequence(batchSize, inputDim);
     }
@@ -240,14 +240,10 @@ TEST(Matrix, unary) {
     // inverse matrix
     testMatrixInverse(height);
 #else
-    LOG(WARNING) << "Cannot run Matrix Inverse Unit Test.\n"
-                 << "Failed to find lapack library in current system.\n"
-                 << "To address this issue, Please adopt one of the following "
-                    "approaches: \n"
-                 << "1. Simply issue `sudo apt-get install liblapacke-dev` to "
-                    "avoid re-build source code. \n"
-                 << "2. Install MKL/Openblas/ATLAS and re-build PaddlePaddle "
-                    "source code.";
+    LOG(WARNING) << "This version of PaddlePaddle was not built with LAPACK"
+                 << "support so we cannot test matrix inverse. To test "
+                 << "matrix inverse, please install LAPACKE "
+                 << "and MKL/Openblas/ATLAS, and re-build PaddlePaddle.";
 #endif
   }
 }
@@ -341,8 +337,8 @@ void testMatrixSoftmaxBp(int height, int width) {
 }
 
 TEST(Matrix, softmax) {
-  for (auto height : {1, 11, 73, 128, 200}) {
-    for (auto width : {1, 32, 100, 512, 1000}) {
+  for (auto height : {1, 3, 131}) {    // prime numbers close to 1, 4, 127
+    for (auto width : {1, 17, 251}) {  // prime numbers close to 1, 16, 256
       VLOG(3) << " height=" << height << " width=" << width;
 
       testMatrixSoftmax(height, width);
@@ -527,7 +523,7 @@ void testVectorRowFunc(int size) {
 }
 
 TEST(Vector, rowFunc) {
-  for (auto size : {1, 5, 31, 90, 150, 500, 1000, 4000}) {
+  for (auto size : {1, 3, 997}) {  // prime numbers close to 1, 4, 1024
     VLOG(3) << " size=" << size;
     testVectorRowFunc(size);
   }
@@ -604,7 +600,7 @@ void testVectorIsEqual(int size) {
 }
 
 TEST(Vector, Equal) {
-  for (auto size : {1, 5, 31, 90, 150, 500, 1000, 4000}) {
+  for (auto size : {1, 3, 997}) {  // prime numbers close to 1, 4, 1024
     VLOG(3) << " size=" << size;
     testVectorReset<int>(size);
     testVectorReset<real>(size);
@@ -635,9 +631,8 @@ void testMatrixTopK(int samples, int dim, int beamSize) {
 }
 
 TEST(Matrix, topK) {
-  for (auto samples : {1, 5, 31, 90, 150, 500}) {
-    for (auto dim :
-         {1, 5, 8, 10, 15, 64, 80, 120, 256, 300, 1280, 5120, 50000}) {
+  for (auto samples : {1, 17, 131}) {  // prime numbers close to 1, 16, 127
+    for (auto dim : {1, 3, 997}) {     // prime numbers close to 1, 4, 1024
       for (auto beamSize : {1, 5, 10, 20, 40, (int)rand() % dim + 1}) {
         if (beamSize > dim) continue;
         VLOG(3) << " samples=" << samples << " beamSize=" << beamSize
@@ -650,6 +645,7 @@ TEST(Matrix, topK) {
 
 void testSMatrixTopK(int samples, int dim, int beamSize, real ratio) {
   int nnz = samples * dim * ratio;
+  if (nnz < 1) nnz = 1;  // Because sparseRand in MathUtil.cpp requires this.
   MatrixPtr cpuSrc = std::make_shared<CpuSparseMatrix>(samples, dim, nnz);
   MatrixPtr gpuSrc = std::make_shared<GpuSparseMatrix>(samples, dim, nnz);
   MatrixPtr cpuVal = std::make_shared<CpuMatrix>(samples, beamSize);
@@ -683,9 +679,9 @@ void testSMatrixTopK(int samples, int dim, int beamSize, real ratio) {
 }
 
 TEST(SMatrix, topK) {
-  for (auto samples : {1, 5, 100}) {
-    for (auto dim : {10000, 10000, 50000}) {
-      for (auto beamSize : {1, 5, 40, 100, 500}) {
+  for (auto samples : {1, 3, 61}) {
+    for (auto dim : {1, 3, 61}) {
+      for (auto beamSize : {1, 3, 61}) {
         for (auto ratio : {0.01, 0.001}) {
           if (beamSize > dim) continue;
           VLOG(3) << " samples=" << samples << " beamSize=" << beamSize
@@ -806,10 +802,9 @@ void testClassificationError(int numSamples, int dim, int topkSize) {
 }
 
 TEST(Matrix, classificationError) {
-  for (auto numSamples : {1, 5, 31, 90, 150, 300}) {
-    for (auto dim :
-         {1, 5, 8, 10, 15, 64, 80, 120, 256, 300, 1280, 5120, 50000}) {
-      for (auto topkSize : {1, 5, 10, 20, 40, (int)rand() % dim + 1}) {
+  for (auto numSamples : {1, 3, 31}) {
+    for (auto dim : {1, 3, 31}) {
+      for (auto topkSize : {1, 3, (int)rand() % dim + 1}) {
         if (topkSize > dim) continue;
         VLOG(3) << " sample= " << numSamples << " topkSize= " << topkSize
                 << " dim= " << dim;
@@ -1016,13 +1011,15 @@ void testAvgPoolFwdBwd(int numSamples,
   TensorCheckErr(*inputGrad, *inputGpuGrad);
 }
 
+// TODO(yi): I noticed many such blindly combinatorial tests in this
+// file.  They are no help to locate defects at all.
 TEST(Matrix, PoolFwdBwd) {
-  for (auto numSamples : {5, 32}) {
-    for (auto channels : {1, 9, 32}) {
-      for (auto imgSizeH : {14, 28}) {
-        for (auto imgSizeW : {16, 30}) {
-          for (auto sizeX : {2, 5}) {
-            for (auto sizeY : {2, 5}) {
+  for (auto numSamples : {1, 3}) {
+    for (auto channels : {1, 3}) {
+      for (auto imgSizeH : {13, 17}) {
+        for (auto imgSizeW : {17, 19}) {
+          for (auto sizeX : {2, 3}) {
+            for (auto sizeY : {2, 3}) {
               for (auto sH : {1, 2}) {
                 for (auto sW : {1, 2}) {
                   for (auto pH : {0, (sizeY - 1) / 2}) {
@@ -1128,8 +1125,8 @@ TEST(Matrix, MaxOutFwdBwd) {
 }
 
 TEST(CpuMatrix, copyFrom) {
-  const size_t height = 1000;
-  const size_t width = 1000;
+  const size_t height = 31;
+  const size_t width = 53;
   CpuMatrix cpu(height, width);
   GpuMatrix gpu(height, width);
   CpuMatrix copy(height, width);
@@ -1149,6 +1146,10 @@ void testBatch2seqPadding(int batchSize, int inputDim) {
 
   IVectorPtr cpuSequence;
   generateSequenceStartPositions(batchSize, cpuSequence);
+  for (int i = 0; i < cpuSequence->getSize(); ++i) {
+    (cpuSequence->getData())[i] += 1;  // so no way that maxSeqLen is 0;
+  }
+
   IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true);
   gpuSequence->copyFrom(*cpuSequence);
 
@@ -1156,45 +1157,46 @@ void testBatch2seqPadding(int batchSize, int inputDim) {
   size_t maxSeqLen = *std::max_element(cpuSequence->getData(),
                                        cpuSequence->getData() + numSeq);
 
+  printf("numSeq = %ld, maxSeqLen = %ld\n", numSeq, maxSeqLen);
   MatrixPtr cBatch = std::make_shared<CpuMatrix>(numSeq * maxSeqLen, inputDim);
   MatrixPtr gBatch = std::make_shared<GpuMatrix>(numSeq * maxSeqLen, inputDim);
   MatrixPtr cCheck = std::make_shared<CpuMatrix>(numSeq * maxSeqLen, inputDim);
 
-  hl_sequence2batch_copy_padding(gBatch->getData(),
-                                 gpuInput->getData(),
-                                 cpuSequence->getData(),
-                                 inputDim,
-                                 maxSeqLen,
-                                 numSeq,
-                                 false,
-                                 true);
-  cCheck->copyFrom(*gBatch);
-
-  int* seqStart = cpuSequence->getData();
-  float* batchData = cBatch->getData();
-  float* seqData = cpuInput->getData();
-  for (size_t i = 0; i < maxSeqLen; i++) {
-    for (size_t j = 0; j < numSeq; j++) {
-      size_t sequenceStart = seqStart[j];
-      size_t sequenceLength = seqStart[j + 1] - seqStart[j];
-      if (i < sequenceLength) {
-        memcpy(batchData + (i * numSeq + j) * inputDim,
-               seqData + (sequenceStart + i) * inputDim,
-               inputDim * sizeof(real));
-      } else {
-        memset(batchData + (i * numSeq + j) * inputDim,
-               0,
-               inputDim * sizeof(real));
-      }
-    }
-  }
-
-  TensorCheckErr(*cBatch, *cCheck);
+  // hl_sequence2batch_copy_padding(gBatch->getData(),
+  //                                gpuInput->getData(),
+  //                                cpuSequence->getData(),
+  //                                inputDim,
+  //                                maxSeqLen,
+  //                                numSeq,
+  //                                false,
+  //                                true);
+  // cCheck->copyFrom(*gBatch);
+
+  // int* seqStart = cpuSequence->getData();
+  // float* batchData = cBatch->getData();
+  // float* seqData = cpuInput->getData();
+  // for (size_t i = 0; i < maxSeqLen; i++) {
+  //   for (size_t j = 0; j < numSeq; j++) {
+  //     size_t sequenceStart = seqStart[j];
+  //     size_t sequenceLength = seqStart[j + 1] - seqStart[j];
+  //     if (i < sequenceLength) {
+  //       memcpy(batchData + (i * numSeq + j) * inputDim,
+  //              seqData + (sequenceStart + i) * inputDim,
+  //              inputDim * sizeof(real));
+  //     } else {
+  //       memset(batchData + (i * numSeq + j) * inputDim,
+  //              0,
+  //              inputDim * sizeof(real));
+  //     }
+  //   }
+  // }
+
+  // TensorCheckErr(*cBatch, *cCheck);
 }
 
 TEST(Matrix, warpCTC) {
-  for (auto batchSize : {51, 526, 2884}) {
-    for (auto inputDim : {32, 512, 2026}) {
+  for (auto batchSize : {1, 3, 17}) {
+    for (auto inputDim : {1, 3, 31}) {
       VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim;
       testBatch2seqPadding(batchSize, inputDim);
     }

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/CMakeLists.txt

@@ -41,6 +41,9 @@ function(op_library TARGET)
     endif()
 endfunction()
 
+cc_library(net_op SRCS net_op.cc DEPS op_registry)
+cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
+
 op_library(add_op SRCS add_op.cc add_op.cu)
 cc_test(add_op_test SRCS add_op_test.cc DEPS add_op)
 
@@ -59,6 +62,6 @@ op_library(sgd_op SRCS sgd_op.cc sgd_op.cu)
 
 op_library(fc_op
     SRCS fc_op.cc
-    DEPS mul_op rowwise_add_op sigmoid_op softmax_op net)
-op_library(recurrent_op SRCS recurrent_op.cc DEPS op_desc tensor op_registry operator net)
+    DEPS mul_op rowwise_add_op sigmoid_op softmax_op net_op)
+op_library(recurrent_op SRCS recurrent_op.cc DEPS op_desc tensor op_registry operator net_op)
 cc_test(recurrent_op_test SRCS recurrent_op_test.cc DEPS recurrent_op gtest mul_op add_op)

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>);
+                       ops::OnehotCrossEntropyOpKernel<ops::GPUPlace, float>);
\ No newline at end of file

paddle/operators/fc_op.cc

@@ -27,7 +27,7 @@ public:
                                {Output("before_act")},
                                {}));
     auto b = Input("b");
-    if (b != EMPTY_VAR_NAME()) {
+    if (b != framework::kEmptyVarName) {
       AddOp(OpRegistry::CreateOp("rowwise_add",
                                  {Output("before_act"), Input("b")},
                                  {Output("before_act")},

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>);
+    paddle::operators::FillZerosLikeKernel<paddle::platform::GPUPlace, float>);
\ No newline at end of file

paddle/operators/mean_op.cc

@@ -41,7 +41,7 @@ public:
 class MeanGradOp : public OperatorWithKernel {
 protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    ctx.Output<Tensor>("X" + GRAD_VAR_SUFFIX())
+    ctx.Output<Tensor>("X" + framework::kGradVarSuffix)
         ->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };

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>);
+REGISTER_OP_GPU_KERNEL(mean_grad, ops::MeanGradKernel<ops::GPUPlace, float>);
\ No newline at end of file

paddle/operators/mean_op.h

@@ -39,10 +39,10 @@ template <typename Place, typename T>
 class MeanGradKernel : public OpKernel {
 public:
   void Compute(const ExecutionContext& context) const override {
-    auto OG = context.Input<Tensor>("Out" + OperatorBase::GRAD_VAR_SUFFIX());
+    auto OG = context.Input<Tensor>("Out" + framework::kGradVarSuffix);
     PADDLE_ENFORCE(framework::product(OG->dims()) == 1,
                    "Mean Gradient should be scalar");
-    auto IG = context.Output<Tensor>("X" + OperatorBase::GRAD_VAR_SUFFIX());
+    auto IG = context.Output<Tensor>("X" + framework::kGradVarSuffix);
     IG->mutable_data<T>(context.GetPlace());
 
     T ig_size = (T)framework::product(IG->dims());

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>);
+REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<ops::GPUPlace, float>);
\ No newline at end of file

paddle/framework/net.cc → paddle/operators/net_op.cc

@@ -14,11 +14,11 @@
   limitations under the License.
 */
 
-#include "paddle/framework/net.h"
+#include "paddle/operators/net_op.h"
 #include "paddle/framework/op_registry.h"
 
 namespace paddle {
-namespace framework {
+namespace operators {
 
 void NetOp::CompleteAddOp(bool calc) {
   add_op_done_ = true;
@@ -74,5 +74,5 @@ std::string NetOp::DebugString() const {
 
 bool NetOp::IsNetOp() const { return true; }
 
-}  // namespace framework
+}  // namespace operators
 }  // namespace paddle

paddle/framework/net.h → paddle/operators/net_op.h

@@ -14,15 +14,17 @@ limitations under the License. */
 
 #pragma once
 
-#include <paddle/framework/op_desc.pb.h>
-#include <paddle/framework/operator.h>
+#include "paddle/framework/op_desc.pb.h"
 #include "paddle/framework/op_proto.pb.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/framework/operator.h"
 #include "paddle/framework/scope.h"
+#include "paddle/operators/type_alias.h"
 #include "paddle/platform/device_context.h"
 
 namespace paddle {
-namespace framework {
+namespace operators {
+
 /**
  * @brief Network is also a type of Operator
  *
@@ -37,13 +39,13 @@ namespace framework {
  * This is the base class of network, all the networks should implement the APIs
  * it defines.
  */
-class NetOp : public OperatorBase {
- public:
+class NetOp : public framework::OperatorBase {
+public:
   /**
    * Infer all the operators' input and output variables' shapes, will be called
    * before every mini-batch
    */
-  void InferShape(const Scope& scope) const override {
+  void InferShape(const framework::Scope& scope) const override {
     for (auto& op : ops_) {
       op->InferShape(scope);
     }
@@ -56,7 +58,7 @@ class NetOp : public OperatorBase {
    * scope will be used instead. If no OpContext is provicded, default context
    * will be used.
    */
-  void Run(const Scope& scope,
+  void Run(const framework::Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {
     for (auto& op : ops_) {
       op->Run(scope, dev_ctx);
@@ -88,7 +90,7 @@ class NetOp : public OperatorBase {
 
   std::vector<std::shared_ptr<OperatorBase>> ops_;
 
- private:
+private:
   bool add_op_done_{false};
 
   template <typename T, typename KeyType>
@@ -97,5 +99,5 @@ class NetOp : public OperatorBase {
   }
 };
 
-}  // namespace framework
+}  // namespace operators
 }  // namespace paddle

paddle/framework/net_op_test.cc → paddle/operators/net_op_test.cc

+#include "paddle/operators/net_op.h"
+
 #include <gtest/gtest.h>
-#include <paddle/framework/net.h>
-#include <paddle/framework/op_registry.h>
-#include <paddle/framework/operator.h>
+
+#include "paddle/framework/op_registry.h"
+#include "paddle/framework/operator.h"
 
 namespace paddle {
-namespace framework {
+namespace operators {
 
 static int infer_shape_cnt = 0;
 static int run_cnt = 0;
 
 class TestOp : public OperatorBase {
- public:
+public:
   void InferShape(const framework::Scope& scope) const override {
     ++infer_shape_cnt;
   }
@@ -21,7 +23,7 @@ class TestOp : public OperatorBase {
 };
 
 class EmptyOp : public OperatorBase {
- public:
+public:
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {}
@@ -73,7 +75,7 @@ TEST(OpKernel, all) {
   ASSERT_THROW(net->AddOp(op2), paddle::platform::EnforceNotMet);
 }
 
-TEST(Net, insert_op) {
+TEST(NetOp, insert_op) {
   NetOp net;
   auto op1 = std::make_shared<EmptyOp>();
   op1->inputs_ = {"x", "w1", "b1"};
@@ -85,5 +87,5 @@ TEST(Net, insert_op) {
   ASSERT_EQ(3UL, net.ops_.size());
 }
 
-}  // namespace framework
+}  // namespace operators
 }  // namespace paddle

paddle/operators/recurrent_op.cc

@@ -18,8 +18,8 @@
 #include <cstring>
 #include <sstream>
 
-#include "paddle/framework/net.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/net_op.h"
 #include "paddle/platform/enforce.h"
 
 namespace paddle {
@@ -38,10 +38,10 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
                    "input link [%s] is not in scope.",
                    inlinks[i].external);
     Tensor* input = input_var->GetMutable<Tensor>();
-    DDim dims = input->dims();
+    framework::DDim dims = input->dims();
     PADDLE_ENFORCE(static_cast<size_t>(dims[0]) == seq_len,
                    "all the inlinks must have same length");
-    DDim step_dims = slice_ddim(dims, 1, dims.size());
+    framework::DDim step_dims = slice_ddim(dims, 1, dims.size());
     for (size_t j = 0; j < seq_len; j++) {
       Tensor* step_input =
           step_scopes[j]->NewVar(inlinks[i].internal)->GetMutable<Tensor>();
@@ -64,13 +64,13 @@ void ConcatOutputs(const std::vector<Scope*>& step_scopes,
                    outlinks[i].external);
     Tensor* output = output_var->GetMutable<Tensor>();
     if (infer_shape_mode) {
-      DDim step_dims = step_scopes[0]
-                           ->FindVar(outlinks[i].internal)
-                           ->GetMutable<Tensor>()
-                           ->dims();
+      framework::DDim step_dims = step_scopes[0]
+                                      ->FindVar(outlinks[i].internal)
+                                      ->GetMutable<Tensor>()
+                                      ->dims();
       std::vector<int> dims_vec = vectorize(step_dims);
       dims_vec.insert(dims_vec.begin(), seq_len);
-      output->Resize(make_ddim(dims_vec));
+      output->Resize(framework::make_ddim(dims_vec));
     } else {
       output->mutable_data<float>(platform::CPUPlace());
       for (size_t j = 0; j < seq_len; j++) {

paddle/operators/recurrent_op.h

@@ -19,8 +19,6 @@
 namespace paddle {
 namespace operators {
 
-using namespace paddle::framework;
-
 namespace rnn {
 
 /**
@@ -70,7 +68,7 @@ struct ArgumentName {
 /**
  * Prepare inputs for each step net.
  */
-void SegmentInputs(const std::vector<Scope*>& step_scopes,
+void SegmentInputs(const std::vector<framework::Scope*>& step_scopes,
                    const std::vector<Link>& inlinks,
                    const size_t seq_len,
                    bool infer_shape_mode);
@@ -78,12 +76,12 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
 /**
  * Process outputs of step nets and merge to variables.
  */
-void ConcatOutputs(const std::vector<Scope*>& step_scopes,
+void ConcatOutputs(const std::vector<framework::Scope*>& step_scopes,
                    const std::vector<Link>& outlinks,
                    const size_t seq_len,
                    bool infer_shape_mode);
 
-void LinkMemories(const std::vector<Scope*>& step_scopes,
+void LinkMemories(const std::vector<framework::Scope*>& step_scopes,
                   const std::vector<MemoryAttr>& memories,
                   const size_t step_id,
                   const int offset,
@@ -94,7 +92,7 @@ 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.
@@ -103,14 +101,15 @@ void InitArgument(const ArgumentName& name, Argument* arg);
 
 class RecurrentAlgorithm {
 public:
-  void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const;
+  void Run(const framework::Scope& scope,
+           const platform::DeviceContext& dev_ctx) const;
 
   void Init(std::unique_ptr<rnn::Argument> arg) { arg_ = std::move(arg); }
 
   /**
    * InferShape must be called before Run.
    */
-  void InferShape(const Scope& scope) const;
+  void InferShape(const framework::Scope& scope) const;
 
 protected:
   /*
@@ -119,13 +118,15 @@ protected:
    * NOTE the scopes are reused in both the forward and backward, so just
    * create once and expand its size if more steps need.
    */
-  void CreateScopes(const Scope& scope) const;
+  void CreateScopes(const framework::Scope& scope) const;
 
-  const std::vector<Scope*>& GetStepScopes(const Scope& scope) const {
-    return *scope.FindVar(arg_->step_scopes)->GetMutable<std::vector<Scope*>>();
+  const std::vector<framework::Scope*>& GetStepScopes(
+      const framework::Scope& scope) const {
+    return *scope.FindVar(arg_->step_scopes)
+                ->GetMutable<std::vector<framework::Scope*>>();
   }
 
-  void InitMemories(Scope* step_scopes, bool infer_shape_mode) const;
+  void InitMemories(framework::Scope* step_scopes, bool infer_shape_mode) const;
 
 private:
   std::unique_ptr<rnn::Argument> arg_;
@@ -146,18 +147,22 @@ class RecurrentGradientAlgorithm {
 public:
   void Init(std::unique_ptr<rnn::Argument> arg) { arg_ = std::move(arg); }
 
-  void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const;
+  void Run(const framework::Scope& scope,
+           const platform::DeviceContext& dev_ctx) const;
 
-  void LinkBootMemoryGradients(Scope* step_scopes, bool infer_shape_mode) const;
+  void LinkBootMemoryGradients(framework::Scope* step_scopes,
+                               bool infer_shape_mode) const;
 
   /**
    * InferShape must be called before Run.
    */
-  void InferShape(const Scope& scope) const;
+  void InferShape(const framework::Scope& scope) const;
 
 protected:
-  inline const std::vector<Scope*>& GetStepScopes(const Scope& scope) const {
-    return *scope.FindVar(arg_->step_scopes)->GetMutable<std::vector<Scope*>>();
+  inline const std::vector<framework::Scope*>& GetStepScopes(
+      const framework::Scope& scope) const {
+    return *scope.FindVar(arg_->step_scopes)
+                ->GetMutable<std::vector<framework::Scope*>>();
   }
 
 private:
@@ -165,19 +170,19 @@ private:
   mutable size_t seq_len_;
 };
 
-class RecurrentOp final : public OperatorBase {
+class RecurrentOp final : public framework::OperatorBase {
 public:
   void Init() override;
 
   /**
    * InferShape must be called before Run.
    */
-  virtual void InferShape(const Scope& scope) const override {
+  void InferShape(const framework::Scope& scope) const override {
     alg_.InferShape(scope);
   }
 
-  virtual void Run(const Scope& scope,
-                   const platform::DeviceContext& dev_ctx) const override {
+  void Run(const framework::Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override {
     alg_.Run(scope, dev_ctx);
   }
 
@@ -187,19 +192,19 @@ private:
   RecurrentAlgorithm alg_;
 };
 
-class RecurrentGradientOp final : public OperatorBase {
+class RecurrentGradientOp final : public framework::OperatorBase {
 public:
   void Init() override;
 
   /**
    * InferShape must be called before Run.
    */
-  virtual void InferShape(const Scope& scope) const override {
+  void InferShape(const framework::Scope& scope) const override {
     alg_.InferShape(scope);
   }
 
-  virtual void Run(const Scope& scope,
-                   const platform::DeviceContext& dev_ctx) const override {
+  void Run(const framework::Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override {
     alg_.Run(scope, dev_ctx);
   }
 

paddle/operators/recurrent_op_test.cc

@@ -11,18 +11,23 @@
   limitations under the License.
 */
 
+#include "paddle/operators/recurrent_op.h"
+
 #include <glog/logging.h>
 #include <gtest/gtest.h>
 
-#include "paddle/framework/net.h"
+#include "paddle/framework/ddim.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/tensor.h"
-#include "paddle/operators/recurrent_op.h"
+#include "paddle/operators/net_op.h"
 
 namespace paddle {
 namespace operators {
 
+using framework::make_ddim;
+using framework::DDim;
+
 class RecurrentOpTest : public ::testing::Test {
 protected:
   virtual void SetUp() override {
@@ -71,7 +76,7 @@ protected:
   }
 
   void CreateRNNOp() {
-    OpDesc op_desc;
+    framework::OpDesc op_desc;
 
     op_desc.set_type("recurrent_op");
     // inlinks 0

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>);
+REGISTER_OP_GPU_KERNEL(sgd, ops::SGDOpKernel<ops::GPUPlace, float>);
\ No newline at end of file

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.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
+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
+    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. */
 
-   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/operators/softmax_op.h"
 
 namespace paddle {
@@ -19,12 +20,13 @@ namespace operators {
 class SoftmaxOp : public OperatorWithKernel {
 protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 1, "Only one input is need for softmax");
-    PADDLE_ENFORCE(ctx.Input<Tensor>(0)->dims().size() == 2,
+    PADDLE_ENFORCE(ctx.InputSize() == 1UL,
+                   "Only one input is need for softmax");
+    PADDLE_ENFORCE(ctx.Input<Tensor>("X")->dims().size() == 2UL,
                    "The input of softmax op must be matrix");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1,
+    PADDLE_ENFORCE(ctx.OutputSize() == 1UL,
                    "Only one output is need for softmax");
-    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
+    ctx.Output<Tensor>("Y")->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
 
@@ -40,10 +42,19 @@ public:
 
 class SoftmaxOpGrad : public OperatorWithKernel {
 protected:
-  void InferShape(const InferShapeContext &ctx) const override {}
-  std::string DebugString() const override {
-    LOG(INFO) << "SoftmaxOpGrad";
-    return "";
+  void InferShape(const InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE(ctx.InputSize() == 3UL,
+                   "Input of SoftmaxOpGrad should be 3, X, Y, YG");
+    PADDLE_ENFORCE(ctx.OutputSize() == 1UL,
+                   "Output of SoftmaxOpGrad should be 1");
+    PADDLE_ENFORCE(ctx.InputVar("Y") != nullptr, "Input(Y) should not be null");
+    PADDLE_ENFORCE(ctx.InputVar(framework::GradVarName("Y")) != nullptr,
+                   "Input(Y@GRAD) should not be null");
+    PADDLE_ENFORCE(ctx.Input<Tensor>("Y")->dims() ==
+                       ctx.Input<Tensor>(framework::GradVarName("Y"))->dims(),
+                   "the shape of Input(0) and Input(1) should be the same");
+    ctx.Output<Tensor>(framework::GradVarName("X"))
+        ->Resize(ctx.Input<Tensor>("Y")->dims());
   }
 };
 
@@ -51,5 +62,7 @@ protected:
 }  // namespace paddle
 
 REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker);
-REGISTER_GRADIENT_OP(softmax, softmax_grad, ops::SoftmaxOpGrad);
 REGISTER_OP_CPU_KERNEL(softmax, ops::SoftmaxKernel<ops::CPUPlace, float>);
+REGISTER_GRADIENT_OP(softmax, softmax_grad, ops::SoftmaxOpGrad);
+REGISTER_OP_CPU_KERNEL(softmax_grad,
+                       ops::SoftmaxGradKernel<ops::CPUPlace, float>);

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>);

paddle/operators/softmax_op.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
+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
+    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. */
+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/ddim.h"
+#include "paddle/framework/operator.h"
+#include "paddle/framework/tensor.h"
 #include "paddle/operators/type_alias.h"
 
 namespace paddle {
@@ -23,8 +26,8 @@ template <typename Place, typename T>
 class SoftmaxKernel : public OpKernel {
 public:
   void Compute(const ExecutionContext& context) const override {
-    auto input = context.Input<Tensor>(0);
-    auto output = context.Output<Tensor>(0);
+    auto input = context.Input<Tensor>("X");
+    auto output = context.Output<Tensor>("Y");
     output->mutable_data<T>(context.GetPlace());
 
     auto logits = EigenMatrix<T>::From(*input);
@@ -57,5 +60,38 @@ public:
              .broadcast(one_by_class));
   }
 };
+
+template <typename Place, typename T>
+class SoftmaxGradKernel : public OpKernel {
+public:
+  void Compute(const ExecutionContext& context) const override {
+    std::shared_ptr<Tensor> scale_ = std::make_shared<Tensor>();
+
+    auto Y = context.Input<Tensor>("Y");
+    auto dY = context.Input<Tensor>(framework::GradVarName("Y"));
+    auto dX = context.Output<Tensor>(framework::GradVarName("X"));
+    dX->mutable_data<T>(context.GetPlace());
+
+    const int batch_size = Y->dims()[0];
+    const int class_num = Y->dims()[1];
+
+    Eigen::DSizes<int, 1> along_class(1);
+    Eigen::DSizes<int, 2> batch_by_one(batch_size, 1);
+    Eigen::DSizes<int, 2> one_by_class(1, class_num);
+
+    auto Y_eigen = EigenMatrix<T>::From(*Y);
+    auto dY_eigen = EigenMatrix<T>::From(*dY);
+    auto dX_eigen = EigenMatrix<T>::From(*dX);
+    auto place = context.GetEigenDevice<Place>();
+
+    auto dot = (Y_eigen * dY_eigen)
+                   .sum(along_class)
+                   .eval()
+                   .reshape(batch_by_one)
+                   .broadcast(one_by_class);
+    dX_eigen.device(place) = (dY_eigen - dot) * Y_eigen;
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle

paddle/operators/type_alias.h

@@ -15,13 +15,14 @@
 #pragma once
 
 #include "paddle/framework/eigen.h"
-#include "paddle/framework/net.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/net_op.h"
 
 namespace paddle {
 namespace operators {
 
 using OpKernel = framework::OpKernel;
+using OperatorBase = framework::OperatorBase;
 using InferShapeContext = framework::InferShapeContext;
 using ExecutionContext = framework::ExecutionContext;
 using Variable = framework::Variable;
@@ -43,15 +44,16 @@ template <typename T,
           typename IndexType = Eigen::DenseIndex>
 using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
 using Tensor = framework::Tensor;
+using Scope = framework::Scope;
 using OperatorWithKernel = framework::OperatorWithKernel;
+using OperatorBase = framework::OperatorBase;
 using OpProtoAndCheckerMaker = framework::OpProtoAndCheckerMaker;
 using OpProto = framework::OpProto;
 using OpAttrChecker = framework::OpAttrChecker;
 using CPUPlace = platform::CPUPlace;
 using GPUPlace = platform::GPUPlace;
-using NetOp = framework::NetOp;
 using OpRegistry = framework::OpRegistry;
-using OperatorBase = framework::OperatorBase;
+
 }  // namespace operators
 }  // namespace paddle
 

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;
@@ -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/device_context_test.cc

@@ -15,24 +15,28 @@ limitations under the License. */
 #include "paddle/platform/device_context.h"
 #include "gtest/gtest.h"
 
-using DEVICE_GPU = Eigen::GpuDevice;
 TEST(Device, Init) {
+  using paddle::platform::DeviceContext;
+  using paddle::platform::CUDADeviceContext;
+  using paddle::platform::GPUPlace;
+
   int count = paddle::platform::GetDeviceCount();
   for (int i = 0; i < count; i++) {
-    paddle::platform::DeviceContext* device_context =
-        new paddle::platform::CUDADeviceContext(i);
+    DeviceContext* device_context = new CUDADeviceContext(GPUPlace(i));
     Eigen::GpuDevice* gpu_device =
-        device_context->template get_eigen_device<DEVICE_GPU>();
+        device_context->template get_eigen_device<Eigen::GpuDevice>();
     ASSERT_NE(nullptr, gpu_device);
     delete device_context;
   }
 }
 
 TEST(Device, CUDADeviceContext) {
+  using paddle::platform::CUDADeviceContext;
+  using paddle::platform::GPUPlace;
+
   int count = paddle::platform::GetDeviceCount();
   for (int i = 0; i < count; i++) {
-    paddle::platform::CUDADeviceContext* device_context =
-        new paddle::platform::CUDADeviceContext(i);
+    CUDADeviceContext* device_context = new CUDADeviceContext(GPUPlace(i));
     Eigen::GpuDevice* gpu_device = device_context->eigen_device();
     ASSERT_NE(nullptr, gpu_device);
     cudnnHandle_t cudnn_handle = device_context->cudnn_handle();

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
+}  // namespace paddle
\ No newline at end of file

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);
-}
-}
-}
+
+}  // namespace dynload
+}  // namespace platform
\ No newline at end of file
+}  // 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) {}
+  explicit GPUPlace(int d) : device(d) {}
 
   // 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: accept C string into Piece.
+  Piece(const std::string& s);  // NOLINT: accept C++ string into Piece.
 
   const char* data() const { return data_; }
   size_t len() const { return size_; }

paddle/trainer/tests/sample_trainer_config_compare_sparse.conf

+#edit-mode: -*- python -*-
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# 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.
+
+#Todo(luotao02) This config is only used for unitest. It is out of date now, and will be updated later.
+
+# Note: when making change to this file, please make sure
+# sample_trainer_config_rnn.conf is changed accordingly so that the uniitest
+# for comparing these two nets can pass (test_CompareTwoNets)
+
+default_initial_std(0.1)
+default_device(0)
+
+word_dim = 999
+l1 = 0
+l2 = 0
+
+model_type("nn")
+
+sparse_update = get_config_arg("sparse_update", bool, False)
+
+TrainData(ProtoData(        
+            type = "proto_sequence",
+            files = ('trainer/tests/train_sparse.list'), 
+            ))
+
+Settings(
+    algorithm='sgd',
+    batch_size=100,
+    learning_rate=0.0001,
+    learning_rate_decay_a=4e-08,
+    learning_rate_decay_b=0.0,
+    learning_rate_schedule='poly',
+)
+
+
+wordvec_dim = 32
+layer2_dim = 16
+layer3_dim = 16
+hidden_dim = 32
+
+slot_names = ["qb", "qw", "tb", "tw"]
+
+def ltr_network(network_name,
+                word_dim=word_dim,
+                wordvec_dim=wordvec_dim,
+                layer2_dim=layer2_dim,
+                layer3_dim=layer3_dim,
+                hidden_dim=hidden_dim,
+                slot_names=slot_names,
+                l1=l1,
+                l2=l2):
+
+    slotnum = len(slot_names)
+    for i in xrange(slotnum):
+        Inputs(slot_names[i] + network_name)
+    for i in xrange(slotnum):
+        Layer(
+            name = slot_names[i] + network_name,
+            type = "data",
+            size = word_dim,
+            device = -1,
+        )
+        Layer(
+            name = slot_names[i] + "_embedding_" + network_name,
+            type = "mixed",
+            size = wordvec_dim,
+            bias = False,
+            device = -1,
+            inputs = TableProjection(slot_names[i] + network_name,
+                                     parameter_name = "embedding.w0",
+                                     decay_rate_l1=l1,
+                                     sparse_remote_update = True,
+                                     sparse_update = sparse_update,
+                                     ),
+        )
+        Layer(
+            name = slot_names[i] + "_rnn1_" + network_name,
+            type = "recurrent",
+            active_type = "tanh",
+            bias = Bias(initial_std = 0,
+                        parameter_name = "rnn1.bias"),
+            inputs = Input(slot_names[i] + "_embedding_" + network_name,
+                           parameter_name = "rnn1.w0")
+        )
+        Layer(
+            name = slot_names[i] + "_rnnlast_" + network_name,
+            type = "seqlastins",
+            inputs = [
+                slot_names[i] + "_rnn1_" + network_name,
+            ],
+        )
+
+    Layer(
+        name = "layer2_" + network_name,
+        type = "fc",
+        active_type = "tanh",
+        size = layer2_dim,
+        bias = Bias(parameter_name = "layer2.bias"),
+        inputs = [Input(slot_name + "_rnnlast_" + network_name, 
+                        parameter_name = "_layer2_" + slot_name + ".w", 
+                        decay_rate = l2, 
+                        initial_smart = True) for slot_name in slot_names]
+    )
+    Layer(
+        name = "layer3_" + network_name,
+        type = "fc",
+        active_type = "tanh",
+        size = layer3_dim,
+        bias = Bias(parameter_name = "layer3.bias"),
+        inputs = [
+            Input("layer2_" + network_name, 
+                  parameter_name = "_layer3.w", 
+                  decay_rate = l2, 
+                  initial_smart = True),
+        ]
+    )
+    Layer(
+        name = "output_" + network_name,
+        type = "fc",
+        size = 1,
+        bias = False,
+        inputs = [
+                  Input("layer3_" + network_name,
+                       parameter_name = "_layerO.w"),
+                 ],
+        )
+
+
+ltr_network("left")
+ltr_network("right")
+Inputs("label")
+Layer(
+    name = "label",
+    type = "data",
+    size = 1,
+    )
+Outputs("cost", "qb_rnnlast_left")
+Layer(
+    name = "cost",
+    type = "rank-cost",
+    inputs = ["output_left", "output_right", "label"],
+    )

paddle/trainer/tests/test_CompareSparse.cpp

@@ -23,7 +23,7 @@ using namespace paddle;  // NOLINT
 using namespace std;     // NOLINT
 
 static const string& configFile1 =
-    "trainer/tests/sample_trainer_config_qb_rnn.conf";
+    "trainer/tests/sample_trainer_config_compare_sparse.conf";
 
 DECLARE_bool(use_gpu);
 DECLARE_string(config);

paddle/trainer/tests/train_sparse.list

+trainer/tests/compare_sparse_data

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()

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

 import unittest
-from op_test_util import OpTestMeta
+
 import numpy as np
+import paddle.v2.framework.core as core
+import paddle.v2.framework.create_op_creation_methods as creation
+
+from op_test_util import OpTestMeta
 
 
 def stable_softmax(x):
@@ -19,5 +23,63 @@ class TestSoftmaxOp(unittest.TestCase):
         self.Y = np.apply_along_axis(stable_softmax, 1, self.X)
 
 
+class TestSoftmaxGradOp(unittest.TestCase):
+    def test_softmax_grad(self):
+        op = creation.op_creations.softmax(X="X", Y="Y")
+        backward_op = core.Operator.backward(op, set())
+        self.assertEqual(backward_op.type(), "softmax_grad")
+        expected = '''Op(softmax_grad), inputs:(X, Y, Y@GRAD), outputs:(X@GRAD).'''
+        self.assertEqual(expected, str(backward_op))
+
+        batch_size = 3
+        class_num = 5
+        # Initialize X and add 1e-2 for numerical stability
+        Y = np.random.rand(batch_size, class_num).astype(np.float32)
+        Y = Y + 1e-2
+        dY = np.random.rand(batch_size, class_num).astype(np.float32)
+
+        # Reference implementation of cross entropy with soft labels
+        def label_softmax_grad(Y, dY):
+            dX = Y * 0.0
+            for i in range(batch_size):
+                d = np.dot(Y[i, :], dY[i, :])
+                dX[i, :] = Y[i, :] * (dY[i, :] - d)
+            return dX
+
+        expected = label_softmax_grad(Y, dY)
+
+        scope = core.Scope()
+        places = []
+        places.append(core.CPUPlace())
+        if core.is_compile_gpu():
+            places.append(core.GPUPlace(0))
+
+        for place in places:
+            y = scope.new_var("Y")
+            y_tensor = y.get_tensor()
+            y_tensor.set_dims([batch_size, class_num])
+            y_tensor.alloc_float(place)
+            y_tensor.set(Y, place)
+
+            dy = scope.new_var("Y@GRAD")
+            dy_tensor = dy.get_tensor()
+            dy_tensor.set_dims([batch_size, class_num])
+            dy_tensor.alloc_float(place)
+            dy_tensor.set(dY, place)
+
+            x = scope.new_var("X")
+            dx = scope.new_var("X@GRAD")
+
+            tensor = scope.find_var("X@GRAD").get_tensor()
+            backward_op.infer_shape(scope)
+            self.assertEqual([batch_size, class_num], tensor.shape())
+
+            ctx = core.DeviceContext.create(place)
+            backward_op.run(scope, ctx)
+            actual = np.array(tensor)
+
+            np.testing.assert_almost_equal(actual, expected, decimal=3)
+
+
 if __name__ == '__main__':
     unittest.main()