Merge branch 'ckpt_m2' of github.com:seiriosPlus/Paddle into ckpt_m2

benchmark/fluid/fluid_benchmark.py

@@ -97,7 +97,7 @@ def dist_transpile(trainer_id, args):
         return train_program, fluid.default_startup_program()
     else:
         raise ValueError(
-            'TRAINING_ROLE environment variable must be either TRAINER or PSERVER'
+            'PADDLE_TRAINING_ROLE environment variable must be either TRAINER or PSERVER'
         )
 
 

benchmark/fluid/kube_gen_job.py

@@ -108,10 +108,10 @@ def gen_job():
     tn_container["ports"][0]["containerPort"] = spreadport
 
     envs.append({"name": "PADDLE_JOB_NAME", "value": args.jobname})
-    envs.append({"name": "TRAINERS", "value": str(args.trainers)})
+    envs.append({"name": "PADDLE_TRAINERS", "value": str(args.trainers)})
     envs.append({"name": "PSERVERS", "value": str(args.pservers)})
     envs.append({"name": "ENTRY", "value": args.entry})
-    envs.append({"name": "PADDLE_INIT_PORT", "value": str(args.port)})
+    envs.append({"name": "PADDLE_PSERVER_PORT", "value": str(args.port)})
     envs.append({"name": "PADDLE_PSERVER_PORT", "value": str(args.port)})
     # NOTE: these directories below are cluster specific, please modify
     # this settings before you run on your own cluster.
@@ -167,16 +167,22 @@ def gen_job():
     tn_container["volumeMounts"] = volumeMounts
 
     ps_container["env"] = envs
-    ps_container["env"].append({"name": "TRAINING_ROLE", "value": "PSERVER"})
+    ps_container["env"].append({
+        "name": "PADDLE_TRAINING_ROLE",
+        "value": "PSERVER"
+    })
     tn_container["env"] = envs
     if args.disttype == "pserver":
         tn_container["env"].append({
-            "name": "TRAINING_ROLE",
+            "name": "PADDLE_TRAINING_ROLE",
             "value": "TRAINER"
         })
     elif args.disttype == "nccl2" or args.disttype == "local":
         # NCCL2 have no training role, set to plain WORKER
-        tn_container["env"].append({"name": "TRAINING_ROLE", "value": "WORKER"})
+        tn_container["env"].append({
+            "name": "PADDLE_TRAINING_ROLE",
+            "value": "WORKER"
+        })
 
     os.mkdir(args.jobname)
     if args.disttype == "pserver":

cmake/external/mkldnn.cmake

@@ -45,7 +45,8 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
 ELSE()
     MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN")
 ENDIF()
-SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-unused-result")
+SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result")
+SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value")
 SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
 SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
 ExternalProject_Add(

doc/fluid/howto/cluster/fluid_cluster_train_cn.md

@@ -168,13 +168,13 @@ cd /paddle/python/paddle/fluid/tests/book
 
 第二步，启动Parameter Server：
 ```bash
-PADDLE_INIT_PORT=6174 PADDLE_INIT_PSERVERS=192.168.1.2 TRAINERS=2 POD_IP=192.168.1.2 PADDLE_INIT_TRAINER_ID=1 TRAINING_ROLE=PSERVER python test_fit_a_line.py
+PADDLE_PSERVER_PORT=6174 PADDLE_PSERVER_IPS=192.168.1.2 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=192.168.1.2 PADDLE_TRAINER_ID=1 PADDLE_TRAINING_ROLE=PSERVER python test_fit_a_line.py
 ```
 执行命令后请等待出现提示： ```Server listening on 192.168.1.2:6174 ```, 表示Paramter Server已经正常启动。
 
 第三步，启动Trainer：
 ```bash
-PADDLE_INIT_PORT=6174 PADDLE_INIT_PSERVERS=192.168.1.3 TRAINERS=2 POD_IP=192.168.1.3 PADDLE_INIT_TRAINER_ID=1 TRAINING_ROLE=TRAINER python test_fit_a_line.py
+PADDLE_PSERVER_PORT=6174 PADDLE_PSERVER_IPS=192.168.1.3 PADDLE_TRAINERS=2 PADDLE_CURRENT_IPP=192.168.1.3 PADDLE_TRAINER_ID=1 PADDLE_TRAINING_ROLE=TRAINER python test_fit_a_line.py
 ```
 由于我们定义的Trainer的数量是2个，因此需要在另外一个计算节点上再启动一个Trainer。
 

doc/fluid/howto/cluster/fluid_recordio.md

@@ -114,8 +114,8 @@ def gen_train_list(file_pattern, trainers, trainer_id):
            ret_list.append(f)
    return ret_list
 
-trainers = int(os.getenv("TRAINERS"))
-trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
+trainers = int(os.getenv("PADDLE_TRAINERS"))
+trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
 data_file = fluid.layers.io.open_files(
     filenames=gen_train_list("./mnist-[0-9]*.recordio", 2, 0),
     thread_num=1,

paddle/contrib/CMakeLists.txt

@@ -14,4 +14,3 @@
 #
 
 add_subdirectory(inference)
-add_subdirectory(tape)

paddle/contrib/tape/README.md

-# Dynamic Graph on Fluid
-
-PaddlePaddle Fluid is targeting the autodiff without tape, which, however, is very
-challenging and we are still way from there. DyNet and PyTorch provide a good design
-idea, the *tape*, that significantly eases the challenge.  Also, DyNet provides
-a C++ API that is as convenient as Python but with higher efficiency and could
-conveniently integrate with industrial/production systems. This package, `tape`,
-combines the good of
-
-1. tape from PyTorch and DyNet
-2. C++ API and core from DyNet
-3. rich set of operators from PaddlePaddle
-
-## Overview
-
-We can implement Dynet-like Tape(See this [survey](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/survey/dynamic_graph.md))
-by wrapping Paddle Fluid's `Operator` and `Variable`.
-
-The user API is straight forward since
-
-1. it is imperative. And it uses host language's control flow logic.
-1. it avoids extra concepts such as `Scope` and `Executor`.
-
-All of these benefits come at the cost of just adding one line `reset_global_tape`
-at every iteration.
-
-## Code Structure
-
-In short, the `Tape` contains a vector of `OpHandle`s. And an `OpHandle` contains its
-`type`, the pointers to the `Variable`s, and necessary attributes.
-
-```c++
-class Variable {
-public:
-  VriableHandle Grad(); // returns its gradient variable
-private:
-  framework::VarDesc desc_; // compile time infershape, necessary for lazy execution
-  framework::Variable var_; // run time variable, holds data memory
-};
-
-using VariableHandle = shared_ptr<Variable>;
-
-struct OpHandle {
-  string type_;
-  map<string, vector<VariableHandle>> inputs_;
-  map<string, vector<VariableHandle>> outputs_;
-  AttributeMap attrs_;
-};
-
-class Tape {
-public:
-  void AddOp(OpHandle); // add op
-  void Forward();       // execute the tape_
-  void Backward();      // execute the backward of the tape_
-private:
-  vector<OpHandle> tape_;
-};
-```
-
-We uses `Function` to indicate layers. It takes care of parameter
-initialization and `AddOp` to the Tape when it is called.
-
-```c++
-class Linear {
- public:
-  Linear(int in_dim, int out_dim, const std::string &act)
-      : w_(new Variable("LinearWeight")),
-        b_(new Variable("LinearBias")),
-        act_(act) {
-    Tape init_tape;
-
-    std::string initializer = "fill_constant";
-    framework::AttributeMap attrs;
-    attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
-    attrs["shape"] = std::vector<int>{in_dim, out_dim};
-    attrs["value"] = 1.0f;
-    init_tape.AddOp(initializer, {}, {{"Out", {w_}}}, attrs);
-
-    attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
-    attrs["shape"] = std::vector<int>{out_dim};
-    attrs["value"] = 1.0f;
-    init_tape.AddOp(initializer, {}, {{"Out", {b_}}}, attrs);
-
-    init_tape.Forward();
-  }
-
-  VariableHandle operator()(VariableHandle input) {
-    VariableHandle pre_bias(new Variable("linear"));
-    get_global_tape().AddOp("mul",
-                            {{"X", {input}}, {"Y", {w_}}},
-                            {{"Out", {pre_bias}}},
-                            {{"x_num_col_dims", 1}, {"y_num_col_dims", 1}});
-    VariableHandle pre_act(new Variable("linear"));
-    get_global_tape().AddOp("elementwise_add",
-                            {{"X", {pre_bias}}, {"Y", {b_}}},
-                            {{"Out", {pre_act}}},
-                            {{"axis", 1}});
-    VariableHandle post_act(new Variable("linear"));
-    get_global_tape().AddOp(act_,
-                            {{"X", {pre_act}}},
-                            {{"Out", {post_act}}},
-                            {});
-    return post_act;
-  }
-
-  std::vector<VariableHandle> Params() { return {w_, b_}; }
-
- private:
-  VariableHandle w_;
-  VariableHandle b_;
-  std::string act_;
-};
-```
-
-## User API
-
-```c++
-// Model function
-paddle::tape::Linear linear1(3, 3, "relu"); // init weight and bias
-paddle::tape::Linear linear2(3, 3, "relu"); // init weight and bias
-paddle::tape::Mean mean;
-
-// Optimizer
-paddle::tape::SGD sgd(0.001);
-
-// Data Feeder
-paddle::tape::Fill data_feeder(...);
-VariableHandle input(new paddle::tape::Variable("input"));
-VariableHandle label(new paddle::tape::Variable("label"));
-
-for (int i = 0; i < 2; ++i) {
-  reset_global_tape();
-
-  data_feeder(input, label);
-
-  auto loss = softmax(linear2(linear1(input)), label); // compile time InferShape & InferVarType
-  LOG(INFO) << loss.value(); // Run forward up to loss
-
-  // Run backward, store gradient of w at w->Grad()
-  get_global_tape.Backward(loss);
-
-  // Update w
-  sgd(linear1.Params());
-  sgd(linear2.Params());
-}
-```
-
-<details>
-  <summary></summary>
-digraph G {
-
-	subgraph cluster_0 {
-                node [shape=record,style=filled];
-		style=filled;
-		color=lightgrey;
-                linear1 [label="{type: mul | {input | {<before_mul1>X: before_mul1 |<weight1> Y: weight1}} |  {output |<before_bias1> Out: before_bias1}}"];
-                elementwise_add1 [label="{type: elementwise_add | {input | {<before_bias1>X: before_bias1 |<bias1> Y: bias1}} |  {output |<before_act1> Out: before_act1}}"];
-                relu1 [label="{type: relu | {input | {<before_act1>X: before_act1 }} |  {output |<after_act1> Out: after_act1}}"];
-
-		linear1 -> elementwise_add1->relu1;
-		label = "forward tape";
-	}
-
-        linear1:before_mul1->before_mul1
-        linear1:weight1->weight1
-        linear1:before_bias1->before_bias1
-
-        elementwise_add1:bias1->bias1
-        elementwise_add1:before_bias1->before_bias1
-        elementwise_add1:before_act1->before_act1
-
-        relu1:before_act1->before_act1
-        relu1:after_act1->after_act1
-
-	subgraph cluster_1 {
-                node [shape=record,style=filled];
-		style=filled;
-		color=lightgrey;
-                linear1_grad [label="{type: mul_grad | {input | {<before_mul1>X: before_mul1 |<weight1> Y: weight1|<before_bias1_grad> Out_grad: before_bias1_grad}} |  {output |{<before_mul1_grad>X_grad: before_mul1_grad |<weight1_grad> Y_grad: weight1_grad}}}"];
-
-                elementwise_add1_grad [label="{type: elementwise_add_grad | {input | <before_act1_grad> Out_grad: before_act1_grad} |  {output |{<before_bias1_grad>X_grad: before_bias1_grad |<bias1_grad> Y_grad: bias1_grad}}}"];
-
-                relu1_grad [label="{type: relu_grad |  {input |<after_act1_grad> Out_grad: after_act1_grad} | {ouput | {<before_act1_grad>X_grad: before_act1_grad }}}"];
-
-		linear1_grad -> elementwise_add1_grad ->relu1_grad [dir=back];
-                label = "backward tape";
-	}
-
-        relu1_grad:after_act1_grad->after_act1_grad
-        relu1_grad:before_act1_grad->before_act1_grad
-
-        elementwise_add1_grad:before_act1_grad->before_act1_grad
-        elementwise_add1_grad:before_bias1_grad->before_bias1_grad
-        elementwise_add1_grad:bias1_grad->bias1_grad
-
-        linear1_grad:before_mul1->before_mul1
-        linear1_grad:weight1->weight1
-        linear1_grad:before_bias1_grad->before_bias1_grad
-        linear1_grad:before_mul1_grad->before_mul1_grad
-        linear1_grad:weight1_grad->weight1_grad
-
-
-	subgraph cluster_2 {
-                node [shape=record];
-                label = "Linear1";
-                weight1
-                bias1
-	}
-
-        weight1 -> weight1_grad [ label="Grad()", style="dashed" ];
-        bias1 -> bias1_grad [ label="Grad()", style="dashed"];
-
-	
-
-}
-</details>
-
-![Image](https://github.com/tonyyang-svail/Paddle/blob/cpp_tap/paddle/contrib/tape/computation_graph.png)
-
-## Code Reuse
-
-We want to stay close to Paddle Fluid as much as possible.
-
-### Reuse All Operators
-
-As all Ops are registered at `OpInfoMap`, the effort of adding a new `Function`
-is about 10 lines of code, similar to expose an operator to Python.
-
-### Reuse Compile Time InferShape and InferVarType
-
-Note that all the symbolic information is stored at `tape::Varaible::desc_`, instead
-of `ProgramDesc.block.vars`, we create a temporary `BlockDesc` to do `InferShape` and
-`InferVarType` every time we `AddOp` to the tape.
-
-### Reuse Operator::Run
-
-We use smart pointer, instead of `Scope`, to manage memory. So we create a temporary
-`Scope` for every `Operator::Run()`.
-
-## Possible Feature
-
-### Release Memory on Backward
-
-We can release memory aggressively. During backward, we can delete the OpHandle once
-we have finished its backward. Since all the variable is managed by smart pointer, the
-memory is automatically released when its `ref_count` goes to 0.
-
-### Kernel Fusion
-
-As a symbolic representation of the Tape is constructed first before the actual
-execution, it would be possible to perform graph optimization. One use case is kernel
-fusion.

paddle/contrib/tape/function.h

-// Copyright (c) 2018 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.
-
-#pragma once
-
-#include <string>
-
-#include "paddle/contrib/tape/tape.h"
-#include "paddle/contrib/tape/variable.h"
-#include "paddle/fluid/framework/type_defs.h"
-
-namespace paddle {
-namespace tape {
-
-class Function {};
-
-class Fill {
- public:
-  Fill(const std::string &initializer, const framework::AttributeMap &attrs)
-      : initializer_(initializer), attrs_(attrs) {}
-
-  void operator()(VariableHandle var) {
-    get_global_tape().AddOp(initializer_, {}, {{"Out", {var}}}, attrs_);
-  }
-
- private:
-  const std::string initializer_;
-  const framework::AttributeMap attrs_;
-};
-
-class Mean {
- public:
-  VariableHandle operator()(VariableHandle var) {
-    VariableHandle out(new Variable("mean"));
-    get_global_tape().AddOp("mean", {{"X", {var}}}, {{"Out", {out}}}, {});
-    return out;
-  }
-};
-
-class Linear {
- public:
-  Linear(int in_dim, int out_dim, const std::string &act)
-      : w_(new Variable("LinearWeight")),
-        b_(new Variable("LinearBias")),
-        act_(act) {
-    Tape init_tape;
-
-    std::string initializer = "fill_constant";
-    framework::AttributeMap attrs;
-    attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
-    attrs["shape"] = std::vector<int>{in_dim, out_dim};
-    attrs["value"] = 1.0f;
-    init_tape.AddOp(initializer, {}, {{"Out", {w_}}}, attrs);
-
-    attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
-    attrs["shape"] = std::vector<int>{out_dim};
-    attrs["value"] = 1.0f;
-    init_tape.AddOp(initializer, {}, {{"Out", {b_}}}, attrs);
-
-    init_tape.Forward();
-  }
-
-  VariableHandle operator()(VariableHandle input) {
-    VariableHandle pre_bias(new Variable("linear"));
-    get_global_tape().AddOp("mul",
-                            {{"X", {input}}, {"Y", {w_}}},
-                            {{"Out", {pre_bias}}},
-                            {{"x_num_col_dims", 1}, {"y_num_col_dims", 1}});
-    VariableHandle pre_act(new Variable("linear"));
-    get_global_tape().AddOp("elementwise_add",
-                            {{"X", {pre_bias}}, {"Y", {b_}}},
-                            {{"Out", {pre_act}}},
-                            {{"axis", 1}});
-    VariableHandle post_act(new Variable("linear"));
-    get_global_tape().AddOp(
-        act_, {{"X", {pre_act}}}, {{"Out", {post_act}}}, {});
-    return post_act;
-  }
-
-  std::vector<VariableHandle> Params() { return {w_, b_}; }
-
- private:
-  VariableHandle w_;
-  VariableHandle b_;
-  std::string act_;
-};
-
-class SGD {
- public:
-  SGD(float learning_rate) : learning_rate_(new Variable("sgd")) {
-    Tape init_tape;
-
-    std::string initializer = "fill_constant";
-    framework::AttributeMap attrs;
-    attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
-    attrs["shape"] = std::vector<int>{1};
-    attrs["value"] = learning_rate;
-    init_tape.AddOp(initializer, {}, {{"Out", {learning_rate_}}}, attrs);
-
-    init_tape.Forward();
-  }
-
-  void operator()(VariableHandle input) {
-    PADDLE_ENFORCE(get_global_tape().HasBeenBackwarded(),
-                   "optimization must happen after the backward");
-    Tape temp_tape;
-    temp_tape.AddOp("sgd",
-                    {{"Param", {input}},
-                     {"LearningRate", {learning_rate_}},
-                     {"Grad", {input->Grad()}}},
-                    {{"ParamOut", {input}}},
-                    {});
-    temp_tape.Forward();
-  }
-
- private:
-  VariableHandle learning_rate_;
-};
-}
-}

paddle/contrib/tape/tape.cc

-// Copyright (c) 2018 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.
-
-#include "paddle/contrib/tape/tape.h"
-
-#include <list>
-#include <map>
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "paddle/fluid/framework/data_type.h"
-#include "paddle/fluid/framework/dim.h"
-#include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/framework/operator.h"
-#include "paddle/fluid/framework/scope.h"
-#include "paddle/fluid/platform/place.h"
-#include "paddle/fluid/pybind/pybind.h"
-
-namespace paddle {
-namespace tape {
-
-// borrowed from
-// https://stackoverflow.com/questions/874134/find-if-string-ends-with-another-string-in-c
-inline bool ends_with(std::string const &value, std::string const &ending) {
-  if (ending.size() > value.size()) return false;
-  return std::equal(ending.rbegin(), ending.rend(), value.rbegin());
-}
-
-std::ostream &operator<<(std::ostream &os, const framework::VarDesc &var_desc) {
-  os << var_desc.Name();
-  os << "[" << var_desc.GetType() << "]";
-  os << "[" << var_desc.GetDataType() << "]";
-  os << "{";
-  for (auto &i : var_desc.GetShape()) {
-    os << i << ",";
-  }
-  os << "}";
-  return os;
-}
-
-std::string to_string(const std::string &type,
-                      const VariableHandleMap &in_vars,
-                      const VariableHandleMap &out_vars,
-                      const framework::AttributeMap &attrs) {
-  std::stringstream ss;
-  ss << type << " ";
-  for (auto &param_name : in_vars) {
-    for (auto &var : param_name.second) {
-      ss << param_name.first << ":(" << var->Desc() << ") ";
-    }
-  }
-  for (auto &param_name : out_vars) {
-    for (auto &var : param_name.second) {
-      ss << param_name.first << ":(" << var->Desc() << ") ";
-    }
-  }
-  return ss.str();
-}
-
-framework::OpDesc CreateOpDesc(const std::string &type,
-                               const VariableHandleMap &in_vars,
-                               const VariableHandleMap &out_vars,
-                               const framework::AttributeMap &attrs) {
-  framework::VariableNameMap inputs;
-  for (auto &param_name : in_vars) {
-    for (auto &var : param_name.second) {
-      inputs[param_name.first].emplace_back(var->Name());
-    }
-  }
-  framework::VariableNameMap outputs;
-  for (auto &param_name : out_vars) {
-    for (auto &var : param_name.second) {
-      outputs[param_name.first].emplace_back(var->Name());
-    }
-  }
-  return framework::OpDesc(type, inputs, outputs, attrs);
-}
-
-void InferShapeAndVarType(const std::string &type,
-                          const VariableHandleMap &in_vars,
-                          VariableHandleMap *out_vars,
-                          const framework::AttributeMap &attrs) {
-  framework::OpDesc op_desc = CreateOpDesc(type, in_vars, *out_vars, attrs);
-
-  // Create a temporary block for compile-time
-  framework::ProgramDesc program_desc;
-  framework::BlockDesc *block_desc = program_desc.MutableBlock(0);
-  PADDLE_ENFORCE(block_desc);
-
-  for (auto &param_name : in_vars) {
-    for (auto &var : param_name.second) {
-      *block_desc->Var(var->Name())->Proto() = *var->MutableDesc()->Proto();
-    }
-  }
-  for (auto &param_name : *out_vars) {
-    for (auto &var : param_name.second) {
-      *block_desc->Var(var->Name())->Proto() = *var->MutableDesc()->Proto();
-    }
-  }
-
-  LOG(INFO) << "- " << to_string(type, in_vars, *out_vars, attrs);
-  op_desc.InferShape(*block_desc);
-  op_desc.InferVarType(block_desc);
-  for (auto &param_name : *out_vars) {
-    for (auto &var : param_name.second) {
-      *var->MutableDesc()->Proto() = *block_desc->Var(var->Name())->Proto();
-    }
-  }
-  LOG(INFO) << "+ " << to_string(type, in_vars, *out_vars, attrs);
-}
-
-void Tape::AddOp(const std::string &type,
-                 const VariableHandleMap &in_vars,
-                 VariableHandleMap out_vars,
-                 const framework::AttributeMap &attrs) {
-  InferShapeAndVarType(type, in_vars, &out_vars, attrs);
-  tape_.emplace_back(type, in_vars, out_vars, attrs);
-}
-
-// Temporary Scope for Operator::Run()
-class ScopeWrapper : public framework::Scope {
- public:
-  ScopeWrapper(const VariableHandleMap &in_vars,
-               const VariableHandleMap &out_vars) {
-    for (auto &v : in_vars) {
-      for (auto &vv : v.second) {
-        if (!vars_.count(vv->Name())) {
-          vars_[vv->Name()].reset(vv->Var());
-        }
-      }
-    }
-    for (auto &v : out_vars) {
-      for (auto &vv : v.second) {
-        if (!vars_.count(vv->Name())) {
-          vars_[vv->Name()].reset(vv->Var());
-        }
-      }
-    }
-  }
-
-  ~ScopeWrapper() {
-    for (auto &pair : vars_) {
-      pair.second.release();
-    }
-  }
-};
-
-void Tape::Forward() {
-  LOG(INFO) << "Starting forward -------------------------";
-  PADDLE_ENFORCE(!has_been_backwarded_);
-  while (current_position_ < tape_.size()) {
-    OpHandle &op = tape_[current_position_];
-
-    // Create Output Tensor, this is only necessary for OpWithKernel
-    for (auto &param2var : op.outputs_) {
-      for (auto &var : param2var.second) {
-        var->InitializeVariable();
-      }
-    }
-
-    framework::OpDesc op_desc =
-        CreateOpDesc(op.type_, op.inputs_, op.outputs_, op.attrs_);
-    ScopeWrapper scope(op.inputs_, op.outputs_);
-    framework::OpRegistry::CreateOp(op_desc)->Run(scope, platform::CPUPlace());
-    current_position_++;
-  }
-
-  LOG(INFO) << "Finishing forward -------------------------";
-}
-
-void Tape::Backward(VariableHandle target) {
-  PADDLE_ENFORCE(!has_been_backwarded_);
-
-  Forward();
-
-  // TODO(tonyyang-svail): check output of last op is target
-  backward_tape_.reset(new Tape());
-
-  framework::AttributeMap attrs;
-
-  // FIXME(tonyyang-svail): Need to infer_data_type
-  attrs["dtype"] = framework::proto::VarType::Type::VarType_Type_FP32;
-  attrs["shape"] = std::vector<int>{1};
-  attrs["value"] = 1.0f;
-  backward_tape_->AddOp(
-      "fill_constant", {}, {{"Out", {target->Grad()}}}, attrs);
-
-  for (auto it = tape_.rbegin(); it != tape_.rend(); ++it) {
-    framework::OpDesc op_desc =
-        CreateOpDesc(it->type_, it->inputs_, it->outputs_, it->attrs_);
-    std::unordered_map<std::string, std::string> grad_to_var;
-    std::vector<std::unique_ptr<framework::OpDesc>> grad_op_descs =
-        framework::OpInfoMap::Instance()
-            .Get(op_desc.Type())
-            .GradOpMaker()(op_desc, {}, &grad_to_var, {});
-
-    for (auto &op_desc : grad_op_descs) {
-      std::unordered_map<std::string, VariableHandle> name2var;
-      for (auto &param2vars : it->inputs_) {
-        for (auto &a : param2vars.second) {
-          name2var[a->Name()] = a;
-        }
-      }
-      for (auto &param2vars : it->outputs_) {
-        for (auto &a : param2vars.second) {
-          name2var[a->Name()] = a;
-        }
-      }
-
-      VariableHandleMap in_vars;
-      VariableHandleMap out_vars;
-      std::map<const framework::VariableNameMap *, VariableHandleMap *>
-          loop_over{{&op_desc->Inputs(), &in_vars},
-                    {&op_desc->Outputs(), &out_vars}};
-      for (auto &each : loop_over) {
-        auto &vmp = *each.first;
-        auto &vhm = *each.second;
-        for (auto &p2a : vmp) {
-          for (auto &argu : p2a.second) {
-            if (name2var.count(argu)) {
-              vhm[p2a.first].push_back(name2var[argu]);
-            } else {
-              PADDLE_ENFORCE(ends_with(argu, framework::kGradVarSuffix),
-                             argu.c_str());
-              std::string name = argu.substr(
-                  0, argu.size() - std::strlen(framework::kGradVarSuffix));
-              PADDLE_ENFORCE(name2var.count(name), name.c_str());
-              vhm[p2a.first].push_back(name2var[name]->Grad());
-            }
-          }
-        }
-      }
-
-      backward_tape_->AddOp(
-          op_desc->Type(), in_vars, out_vars, op_desc->GetAttrMap());
-    }
-
-    // TODO(tonyyang-svail): how to fill empty grad?
-    // TODO(tonyyang-svail): Sum var grad is necessary
-  }
-
-  backward_tape_->Forward();
-  has_been_backwarded_ = true;
-}
-
-Tape &get_global_tape() {
-  static Tape T;
-  return T;
-}
-
-void reset_global_tape() { get_global_tape() = Tape(); }
-}
-}

paddle/contrib/tape/tape.h

-// Copyright (c) 2018 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.
-#pragma once
-
-#include <map>
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "paddle/contrib/tape/variable.h"
-
-namespace paddle {
-namespace tape {
-
-using VariableHandleMap = std::map<std::string, std::vector<VariableHandle>>;
-
-struct OpHandle {
-  OpHandle(const std::string &type,
-           const VariableHandleMap &in_vars,
-           const VariableHandleMap &out_vars,
-           const framework::AttributeMap &attrs)
-      : type_(type), inputs_(in_vars), outputs_(out_vars), attrs_(attrs) {}
-
-  std::string type_;
-  VariableHandleMap inputs_;
-  VariableHandleMap outputs_;
-  framework::AttributeMap attrs_;
-};
-
-class Tape {
- public:
-  void AddOp(const std::string &type,
-             const VariableHandleMap &in_vars,
-             VariableHandleMap out_vars,
-             const framework::AttributeMap &attrs);
-  void Forward();
-  void Backward(VariableHandle target);
-
-  bool HasBeenBackwarded() { return has_been_backwarded_; }
-
- private:
-  bool has_been_backwarded_ = false;
-  size_t current_position_ = 0;
-
-  std::vector<OpHandle> tape_;
-  std::shared_ptr<Tape> backward_tape_;
-};
-
-Tape &get_global_tape();
-
-void reset_global_tape();
-}
-}

paddle/contrib/tape/test_tape.cc

-// Copyright (c) 2018 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.
-
-#include "gtest/gtest.h"
-#include "paddle/contrib/tape/function.h"
-
-using namespace paddle::tape;
-
-TEST(Tape, TestMLP) {
-  LOG(INFO) << "TestMLP";
-  Linear linear1(3, 3, "relu");
-  Linear linear2(3, 3, "relu");
-  Mean mean;
-
-  SGD sgd(0.001);
-
-  std::string initializer = "fill_constant";
-  paddle::framework::AttributeMap attrs;
-  attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
-  attrs["shape"] = std::vector<int>{3, 3};
-  attrs["value"] = 1.0f;
-  Fill filler(initializer, attrs);
-
-  for (int i = 0; i < 2; ++i) {
-    reset_global_tape();
-
-    VariableHandle input(new Variable("input"));
-    filler(input);
-
-    auto loss = mean(linear2(linear1(input)));
-
-    get_global_tape().Backward(loss);
-
-    for (auto w : linear1.Params()) {
-      sgd(w);
-    }
-    for (auto w : linear2.Params()) {
-      sgd(w);
-    }
-  }
-}
-
-int main(int argc, char** argv) {
-  std::vector<paddle::platform::Place> places;
-  places.emplace_back(paddle::platform::CPUPlace());
-  paddle::platform::DeviceContextPool::Init(places);
-
-  testing::InitGoogleTest(&argc, argv);
-  return RUN_ALL_TESTS();
-}

paddle/contrib/tape/variable.cc

-// Copyright (c) 2018 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.
-
-#include "paddle/contrib/tape/variable.h"
-
-namespace paddle {
-namespace tape {
-
-void Variable::InitializeVariable() {
-  LOG(INFO) << "Initialzing " << desc_.Name() << " as " << desc_.GetType();
-  framework::proto::VarType::Type var_type = desc_.GetType();
-  if (var_type == framework::proto::VarType::LOD_TENSOR) {
-    var_.GetMutable<framework::LoDTensor>();
-  } else if (var_type == framework::proto::VarType::SELECTED_ROWS) {
-    var_.GetMutable<framework::SelectedRows>();
-  } else {
-    PADDLE_THROW("Variable type %d is not in [LOD_TENSOR, SELECTED_ROWS]",
-                 var_type);
-  }
-}
-}
-}

paddle/contrib/tape/variable.h

-// Copyright (c) 2018 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.
-#pragma once
-
-#include <memory>
-
-#include "paddle/fluid/framework/operator.h"  // framework::kGradVarSuffix
-#include "paddle/fluid/framework/program_desc.h"
-#include "paddle/fluid/framework/variable.h"
-
-namespace paddle {
-namespace tape {
-
-class Variable;
-using VariableHandle = std::shared_ptr<Variable>;
-
-/*
- * Combination of
- *     framework::VarDesc desc_;
- *     framework::Variable var_;
- */
-class Variable {
- public:
-  Variable(const std::string pre_fix)
-      : desc_(pre_fix + std::to_string(count())) {}
-
-  Variable(const std::string pre_fix, bool is_grad)
-      : desc_(pre_fix + (is_grad ? framework::kGradVarSuffix
-                                 : std::to_string(count()))) {}
-
-  ~Variable() { LOG(INFO) << "Deleting " << Name(); }
-
-  // Instantiate LoDTensor/SelectedRow
-  void InitializeVariable();
-
-  VariableHandle Grad() {
-    if (grad_.expired()) {
-      VariableHandle new_grad(new Variable(desc_.Name(), true));
-      grad_ = new_grad;
-      return new_grad;
-    } else {
-      return VariableHandle(grad_);
-    }
-  }
-
-  // Stochastic Gradient Descent with Momentum
-  //  VariableHandle Momentum ();
-
-  //  void init(const std::string& initializer,
-  //            const framework::AttributeMap& attrs);
-
-  // void value() {};
-
-  const framework::VarDesc& Desc() const { return desc_; }
-  framework::VarDesc* MutableDesc() { return &desc_; }
-
-  // TODO(tonyyang-svail): No need to expose name
-  std::string Name() const { return desc_.Name(); }
-
-  framework::Variable* Var() { return &var_; }
-
- private:
-  int count() {
-    static int counter = 0;
-    return counter++;
-  }
-
-  framework::VarDesc desc_;
-  framework::Variable var_;
-
-  std::weak_ptr<Variable> grad_;
-};
-}
-}

paddle/fluid/framework/executor.cc

@@ -295,13 +295,14 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
 
 std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
     const ProgramDesc& program, int block_id) {
-  auto* ctx = new ExecutorPrepareContext(program, block_id);
+  std::unique_ptr<ExecutorPrepareContext> ctx(
+      new ExecutorPrepareContext(program, block_id));
   PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), program.Size());
   auto& block = program.Block(block_id);
   for (auto& op_desc : block.AllOps()) {
     ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
   }
-  return std::unique_ptr<ExecutorPrepareContext>(ctx);
+  return ctx;
 }
 
 std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(

paddle/fluid/inference/analysis/argument.h

@@ -21,6 +21,8 @@
  * big.
  */
 
+#pragma once
+
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/inference/analysis/data_flow_graph.h"
 
@@ -43,7 +45,7 @@ struct Argument {
 
 #define UNLIKELY(condition) __builtin_expect(static_cast<bool>(condition), 0)
 #define ANALYSIS_ARGUMENT_CHECK_FIELD(field__)               \
-  if (!UNLIKELY(field__)) {                                  \
+  if (UNLIKELY(!(field__))) {                                \
     LOG(ERROR) << "field " << #field__ << " should be set."; \
     return false;                                            \
   }

paddle/fluid/operators/activation_mkldnn_op.cc

@@ -12,16 +12,20 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#include "mkldnn.hpp"
 #include "paddle/fluid/operators/activation_op.h"
-#include "paddle/fluid/operators/mkldnn_activation_op.h"
 #include "paddle/fluid/platform/mkldnn_helper.h"
 
 namespace paddle {
 namespace operators {
 
-using paddle::framework::Tensor;
-using paddle::platform::MKLDNNDeviceContext;
+using framework::DataLayout;
+using framework::Tensor;
+using mkldnn::memory;
+using mkldnn::primitive;
+using mkldnn::stream;
+using platform::GetMKLDNNFormat;
+using platform::MKLDNNDeviceContext;
+using platform::to_void_cast;
 
 namespace {
 std::string gethash(const mkldnn::memory::dims &operand_dims,
@@ -35,188 +39,260 @@ std::string gethash(const mkldnn::memory::dims &operand_dims,
   };
   return dim2str(operand_dims) + std::to_string(algorithm);
 }
+}  // namespace
+
+template <typename Functor>
+class MKLDNNActivationKernel
+    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    const auto *x = ctx.Input<Tensor>("X");
+    PADDLE_ENFORCE(x->layout() == DataLayout::kMKLDNN &&
+                       x->format() != memory::format::format_undef,
+                   "Wrong layout/format set for Input x tensor");
+
+    Functor functor;
+
+    auto attrs = functor.GetAttrs();
+    for (auto &attr : attrs) {
+      *attr.second = ctx.Attr<float>(attr.first);
+    }
+    functor(ctx);
+  }
+};
 
-template <typename T, typename ExecContext>
-void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm,
-                     const T alpha = 0, const T beta = 0) {
+template <typename Functor>
+class MKLDNNActivationGradKernel
+    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    const auto *diff_y = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    PADDLE_ENFORCE(diff_y->layout() == DataLayout::kMKLDNN &&
+                       diff_y->format() != memory::format::format_undef,
+                   "Wrong layout/format set for Input OutGrad tensor");
+
+    Functor functor;
+
+    auto attrs = functor.GetAttrs();
+    for (auto &attr : attrs) {
+      *attr.second = ctx.Attr<float>(attr.first);
+    }
+    functor(ctx);
+  }
+};
+
+template <typename T>
+void eltwise_forward(const framework::ExecutionContext &ctx,
+                     mkldnn::algorithm algorithm, const T alpha = 0,
+                     const T beta = 0) {
   PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
                  "It must use CPUPlace.");
-
   auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
   const auto &mkldnn_engine = dev_ctx.GetEngine();
 
-  // get buffers
-  const auto *src = ctx.template Input<Tensor>("X");
-  const auto *src_data = src->template data<T>();
+  const auto *x = ctx.Input<Tensor>("X");
+  auto *y = ctx.Output<Tensor>("Out");
 
-  auto *dst = ctx.template Output<Tensor>("Out");
-  T *dst_data = dst->template mutable_data<T>(ctx.GetPlace());
+  const T *x_data = x->data<T>();
+  T *y_data = y->mutable_data<T>(ctx.GetPlace());
 
-  // get memory dim
-  PADDLE_ENFORCE(src->dims().size() == 2 || src->dims().size() == 4,
+  PADDLE_ENFORCE(x->dims().size() == 2 || x->dims().size() == 4,
                  "Input dim must be with 2 or 4");
-  std::vector<int> src_tz = framework::vectorize2int(src->dims());
+
+  std::vector<int> src_tz = framework::vectorize2int(x->dims());
+
+  auto src_format =
+      src_tz.size() == 2 ? mkldnn::memory::format::nc : x->format();
 
   const std::string key = gethash(src_tz, algorithm);
   const std::string key_src_data =
       key + ctx.op().Output("Out") + "@eltwise_fwd_src_data";
-  const std::string key_src_mem = key + "@eltwise_fwd_src_mem";
-  const std::string key_dst_mem = key + "@eltwise_fwd_dst_mem";
-  const std::string key_fwd = key + "@eltwise_fwd";
+  const std::string key_src_layout =
+      key + ctx.op().Output("Out") + "@eltwise_fwd_src_layout";
+  const std::string key_with_layout = key + std::to_string(src_format);
+  const std::string key_src_mem = key_with_layout + "@eltwise_fwd_src_mem";
+  const std::string key_dst_mem = key_with_layout + "@eltwise_fwd_dst_mem";
+  const std::string key_fwd = key_with_layout + "@eltwise_fwd";
+  const std::string key_fwd_pd = key_with_layout + "@eltwise_fwd_pd";
+
+  // save input data and layout to be referred in backward path
+  auto p_src_data = std::make_shared<const T *>(x_data);
+  dev_ctx.SetBlob(key_src_data, p_src_data);
+  auto p_src_layout = std::make_shared<memory::format>(src_format);
+  dev_ctx.SetBlob(key_src_layout, p_src_layout);
 
   auto p_fwd = std::static_pointer_cast<mkldnn::eltwise_forward>(
       dev_ctx.GetBlob(key_fwd));
 
-  // save input data to be referred in backward path
-  auto p_src_data = std::make_shared<const T *>(src_data);
-  dev_ctx.SetBlob(key_src_data, p_src_data);
+  std::shared_ptr<memory> dst_memory;
 
   if (p_fwd == nullptr) {
-    // create memory description
-    auto data_md = src_tz.size() == 2
-                       ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
-                                                 mkldnn::memory::format::nc)
-                       : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
-                                                 mkldnn::memory::format::nchw);
-
-    // create memory primitives
-    auto p_src_mem = std::make_shared<mkldnn::memory>(mkldnn::memory(
-        {data_md, mkldnn_engine}, platform::to_void_cast(src_data)));
-    dev_ctx.SetBlob(key_src_mem, p_src_mem);
-
-    auto p_dst_mem = std::make_shared<mkldnn::memory>(mkldnn::memory(
-        {data_md, mkldnn_engine}, platform::to_void_cast(dst_data)));
-    dev_ctx.SetBlob(key_dst_mem, p_dst_mem);
-
-    auto fwd_desc = mkldnn::eltwise_forward::desc(
-        mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta);
-    auto p_fwd_pd = std::make_shared<mkldnn::eltwise_forward::primitive_desc>(
-        fwd_desc, mkldnn_engine);
-    const std::string key_fwd_pd = key + "eltwise_fwd_pd";
-    dev_ctx.SetBlob(key_fwd_pd, p_fwd_pd);
-    p_fwd = std::make_shared<mkldnn::eltwise_forward>(
-        *p_fwd_pd, *(p_src_mem.get()), *(p_dst_mem.get()));
+    // create mkldnn memory for input X
+    auto src_md = platform::MKLDNNMemDesc(
+        src_tz, platform::MKLDNNGetDataType<T>(), src_format);
+    auto src_memory = std::shared_ptr<memory>(
+        new memory({src_md, mkldnn_engine}, to_void_cast(x_data)));
+    // save src_memory to be referred in backward path
+    dev_ctx.SetBlob(key_src_mem, src_memory);
+
+    // create primitive descriptor for activation forward and save it
+    auto forward_desc = mkldnn::eltwise_forward::desc(
+        mkldnn::prop_kind::forward_training, algorithm,
+        src_memory->get_primitive_desc().desc(), alpha, beta);
+    auto forward_pd = std::make_shared<mkldnn::eltwise_forward::primitive_desc>(
+        forward_desc, mkldnn_engine);
+
+    // save prim desc into global device context to be referred in backward path
+    dev_ctx.SetBlob(key_fwd_pd, forward_pd);
+
+    // create mkldnn memory for output y
+    dst_memory =
+        std::make_shared<memory>(forward_pd->dst_primitive_desc(), y_data);
+
+    dev_ctx.SetBlob(key_dst_mem, dst_memory);
+
+    // create activation primitive
+    p_fwd = std::make_shared<mkldnn::eltwise_forward>(*forward_pd, *src_memory,
+                                                      *dst_memory);
     dev_ctx.SetBlob(key_fwd, p_fwd);
   } else {
     // primitives already exist
-    auto p_src_mem =
+    auto src_memory =
         std::static_pointer_cast<mkldnn::memory>(dev_ctx.GetBlob(key_src_mem));
-    PADDLE_ENFORCE(p_src_mem != nullptr,
-                   "Fail to find eltwise p_src_mem in device context.");
-    auto p_dst_mem =
+    PADDLE_ENFORCE(src_memory != nullptr,
+                   "Fail to find eltwise src_memory in device context.");
+    dst_memory =
         std::static_pointer_cast<mkldnn::memory>(dev_ctx.GetBlob(key_dst_mem));
-    PADDLE_ENFORCE(p_dst_mem != nullptr,
-                   "Fail to find eltwise p_src_mem in device context.");
+    PADDLE_ENFORCE(dst_memory != nullptr,
+                   "Fail to find eltwise dst_memory in device context.");
 
-    p_src_mem->set_data_handle(platform::to_void_reinterpret_cast(src_data));
-    p_dst_mem->set_data_handle(dst_data);
+    src_memory->set_data_handle(platform::to_void_cast(x_data));
+    dst_memory->set_data_handle(y_data);
   }
 
   // push primitive to stream and wait until it's executed
-  std::vector<mkldnn::primitive> pipeline = {*(p_fwd.get())};
-  mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
+  std::vector<primitive> pipeline;
+  pipeline.push_back(*p_fwd);
+  stream(stream::kind::eager).submit(pipeline).wait();
+
+  y->set_layout(DataLayout::kMKLDNN);
+  y->set_format(GetMKLDNNFormat(*dst_memory));
 }
 
-template <typename T, typename ExecContext>
-void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm,
-                  const T alpha = 0, const T beta = 0) {
+template <typename T>
+void eltwise_grad(const framework::ExecutionContext &ctx,
+                  mkldnn::algorithm algorithm, const T alpha = 0,
+                  const T beta = 0) {
   auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
   const auto &mkldnn_engine = dev_ctx.GetEngine();
 
-  // get buffers
-  const auto *out = ctx.template Input<Tensor>("Out");
-
-  auto *dout = ctx.template Input<Tensor>(framework::GradVarName("Out"));
-  const auto *diff_dst = dout->template data<T>();
+  const auto *diff_y = ctx.Input<Tensor>(framework::GradVarName("Out"));
+  auto *diff_x = ctx.Output<Tensor>(framework::GradVarName("X"));
 
-  auto *dx =
-      ctx.template Output<framework::Tensor>(framework::GradVarName("X"));
-  const T *diff_src = dx->template mutable_data<T>(ctx.GetPlace());
+  const T *diff_y_data = diff_y->data<T>();
+  T *diff_x_data = diff_x->mutable_data<T>(ctx.GetPlace());
 
-  // get memory dim
-  std::vector<int> src_tz = framework::vectorize2int(out->dims());
+  std::vector<int> diff_dst_tz = framework::vectorize2int(diff_y->dims());
 
-  const std::string key = gethash(src_tz, algorithm);
-  const std::string key_diff_src_mem = key + "@eltwise_diff_src_mem";
-  const std::string key_diff_dst_mem = key + "@eltwise_diff_dst_mem";
-  const std::string key_grad = key + "@eltwise_grad";
+  auto diff_y_format =
+      diff_dst_tz.size() == 2 ? mkldnn::memory::format::nc : diff_y->format();
 
+  const std::string key = gethash(diff_dst_tz, algorithm);
   const std::string key_src_data =
       key + ctx.op().Input("Out") + "@eltwise_fwd_src_data";
+  const std::string key_src_layout =
+      key + ctx.op().Input("Out") + "@eltwise_fwd_src_layout";
+  const auto p_src_layout =
+      std::static_pointer_cast<memory::format>(dev_ctx.GetBlob(key_src_layout));
+  const std::string key_src_mem =
+      key + std::to_string(*p_src_layout) + "@eltwise_fwd_src_mem";
+  const std::string key_fwd_pd =
+      key + std::to_string(*p_src_layout) + "@eltwise_fwd_pd";
+  const std::string key_with_layouts =
+      key + std::to_string(*p_src_layout) + "-" + std::to_string(diff_y_format);
+  const std::string key_diff_src_mem =
+      key_with_layouts + "@eltwise_diff_src_mem";
+  const std::string key_diff_dst_mem =
+      key_with_layouts + "@eltwise_diff_dst_mem";
+  const std::string key_grad = key_with_layouts + "@eltwise_grad";
+
   const auto p_src_data =
       std::static_pointer_cast<T *>(dev_ctx.GetBlob(key_src_data));
 
-  const std::string key_src_mem = key + "@eltwise_fwd_src_mem";
-  auto p_src_mem =
+  auto src_memory =
       std::static_pointer_cast<mkldnn::memory>(dev_ctx.GetBlob(key_src_mem));
-  p_src_mem->set_data_handle(*p_src_data.get());
+  PADDLE_ENFORCE(src_memory != nullptr,
+                 "Fail to find src_memory in device context");
+  src_memory->set_data_handle(*p_src_data.get());
+
+  std::shared_ptr<memory> diff_src_memory;
 
-  auto p_grad = std::static_pointer_cast<mkldnn::eltwise_forward::primitive>(
+  auto p_grad = std::static_pointer_cast<mkldnn::eltwise_backward>(
       dev_ctx.GetBlob(key_grad));
 
   if (p_grad == nullptr) {
-    // create memory description
-    auto data_md = src_tz.size() == 2
-                       ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
-                                                 mkldnn::memory::format::nc)
-                       : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
-                                                 mkldnn::memory::format::nchw);
-
-    // create memory primitives
-    std::shared_ptr<void> p_diff_src_mem =
-        std::make_shared<mkldnn::memory>(mkldnn::memory(
-            {data_md, mkldnn_engine}, platform::to_void_cast(diff_src)));
-    dev_ctx.SetBlob(key_diff_src_mem, p_diff_src_mem);
-    std::shared_ptr<void> p_diff_dst_mem =
-        std::make_shared<mkldnn::memory>(mkldnn::memory(
-            {data_md, mkldnn_engine}, platform::to_void_cast(diff_dst)));
-    dev_ctx.SetBlob(key_diff_dst_mem, p_diff_dst_mem);
-
-    auto bwd_desc = mkldnn::eltwise_backward::desc(algorithm, data_md, data_md,
-                                                   alpha, beta);
-
-    const std::string key_fwd_pd = key + "eltwise_fwd_pd";
-    auto *p_fwd_pd = static_cast<mkldnn::eltwise_forward::primitive_desc *>(
-        dev_ctx.GetBlob(key_fwd_pd).get());
-
-    auto eltwise_bwd_prim_desc = mkldnn::eltwise_backward::primitive_desc(
-        bwd_desc, mkldnn_engine, *p_fwd_pd);
-
+    // create mkldnn memory for input diff_y
+    auto diff_dst_md = platform::MKLDNNMemDesc(
+        diff_dst_tz, platform::MKLDNNGetDataType<T>(), diff_y_format);
+    auto diff_dst_memory = std::shared_ptr<memory>(
+        new memory({diff_dst_md, mkldnn_engine}, to_void_cast(diff_y_data)));
+    dev_ctx.SetBlob(key_diff_dst_mem, diff_dst_memory);
+
+    // retrieve eltwise primitive desc from device context
+    auto forward_pd =
+        std::static_pointer_cast<mkldnn::eltwise_forward::primitive_desc>(
+            dev_ctx.GetBlob(key_fwd_pd));
+    PADDLE_ENFORCE(forward_pd != nullptr,
+                   "Fail to find eltwise_fwd_pd in device context");
+
+    // ceate primitive descriptor for activation backward
+    auto backward_desc = mkldnn::eltwise_backward::desc(
+        algorithm, diff_dst_memory->get_primitive_desc().desc(),
+        src_memory->get_primitive_desc().desc(), alpha, beta);
+    auto backward_pd = mkldnn::eltwise_backward::primitive_desc(
+        backward_desc, mkldnn_engine, *forward_pd);
+
+    // create mkldnn memory for output diff_src
+    diff_src_memory = std::make_shared<memory>(
+        backward_pd.diff_src_primitive_desc(), diff_x_data);
+    dev_ctx.SetBlob(key_diff_src_mem, diff_src_memory);
+
+    // create activation backward primitive
     p_grad = std::make_shared<mkldnn::eltwise_backward>(
-        eltwise_bwd_prim_desc, *static_cast<mkldnn::memory *>(p_src_mem.get()),
-        *(static_cast<mkldnn::memory *>(p_diff_dst_mem.get())),
-        *(static_cast<mkldnn::memory *>(p_diff_src_mem.get())));
+        backward_pd, *src_memory, *diff_dst_memory, *diff_src_memory);
+    dev_ctx.SetBlob(key_grad, p_grad);
   } else {
     // primitives already exist
-    auto p_diff_src_mem = std::static_pointer_cast<mkldnn::memory>(
+    diff_src_memory = std::static_pointer_cast<mkldnn::memory>(
         dev_ctx.GetBlob(key_diff_src_mem));
-    auto p_diff_dst_mem = std::static_pointer_cast<mkldnn::memory>(
+    auto diff_dst_memory = std::static_pointer_cast<mkldnn::memory>(
         dev_ctx.GetBlob(key_diff_dst_mem));
 
-    p_diff_src_mem->set_data_handle(
-        platform::to_void_reinterpret_cast(diff_src));
-    p_diff_dst_mem->set_data_handle(
-        platform::to_void_reinterpret_cast(diff_dst));
+    diff_src_memory->set_data_handle(
+        platform::to_void_reinterpret_cast(diff_x_data));
+    diff_dst_memory->set_data_handle(
+        platform::to_void_reinterpret_cast(diff_y_data));
   }
 
   // push primitive to stream and wait until it's executed
-  std::vector<mkldnn::primitive> pipeline = {*(p_grad.get())};
-  mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
+  std::vector<primitive> pipeline;
+  pipeline.push_back(*p_grad);
+  stream(stream::kind::eager).submit(pipeline).wait();
+
+  diff_x->set_layout(DataLayout::kMKLDNN);
+  diff_x->set_format(GetMKLDNNFormat(*diff_src_memory));
 }
-}  // anonymous namespace
 
 template <typename T, mkldnn::algorithm algorithm>
 struct MKLDNNActivationFunc : public BaseActivationFunctor<T> {
-  template <typename ExecContext>
-  void operator()(const ExecContext &ctx) const {
+  void operator()(const framework::ExecutionContext &ctx) const {
     eltwise_forward<T>(ctx, algorithm);
   }
 };
 
 template <typename T, mkldnn::algorithm algorithm>
 struct MKLDNNActivationGradFunc : public BaseActivationFunctor<T> {
-  template <typename ExecContext>
-  void operator()(const ExecContext &ctx) const {
+  void operator()(const framework::ExecutionContext &ctx) const {
     eltwise_grad<T>(ctx, algorithm);
   }
 };

paddle/fluid/operators/activation_op.cc

@@ -19,6 +19,8 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+using paddle::framework::Tensor;
+
 #define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT)               \
   class OP_NAME##OpMaker                                                \
       : public ::paddle::framework::OpProtoAndCheckerMaker {            \
@@ -27,9 +29,9 @@ namespace operators {
       AddInput("X", "Input of " #OP_NAME " operator");                  \
       AddOutput("Out", "Output of " #OP_NAME " operator").Reuse("X");   \
       AddAttr<bool>("use_mkldnn",                                       \
-                    "(default false) Only used in mkldnn kernel")     \
+                    "(bool, default false) Only used in mkldnn kernel") \
           .SetDefault(false);                                           \
-      AddComment(OP_COMMENT);                                         \
+      AddComment(#OP_COMMENT);                                          \
     }                                                                   \
   }
 
@@ -58,7 +60,6 @@ framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx,
                                       const framework::OperatorWithKernel& oper,
                                       const std::string& name) {
   framework::LibraryType library{framework::LibraryType::kPlain};
-
   framework::DataLayout layout = framework::DataLayout::kAnyLayout;
 #ifdef PADDLE_WITH_MKLDNN
   auto it = oper.Attrs().find("use_mkldnn");
@@ -82,6 +83,7 @@ class ActivationOp : public framework::OperatorWithKernel {
     ctx->ShareLoD("X", /*->*/ "Out");
   }
 
+ protected:
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return GetKernelType(ctx, *this, "X");
@@ -96,6 +98,7 @@ class ActivationOpGrad : public framework::OperatorWithKernel {
     ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out"));
   }
 
+ protected:
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return GetKernelType(ctx, *this, "Out");
@@ -140,7 +143,7 @@ $$out = \\frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$
 __attribute__((unused)) constexpr char TanhShrinkDoc[] = R"DOC(
 TanhShrink Activation Operator.
 
-$$out = x - \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$
+$$out = x - \\frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$
 
 )DOC";
 
@@ -382,7 +385,7 @@ class STanhOpMaker : public framework::OpProtoAndCheckerMaker {
     AddComment(R"DOC(
 STanh Activation Operator.
 
-$$out = b * \frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$
+$$out = b * \\frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$
 
 )DOC");
   }

paddle/fluid/operators/concat_op.h

@@ -60,34 +60,45 @@ template <typename DeviceContext, typename T>
 class ConcatGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
-    auto* in = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto* out_grad =
+        ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto ins = ctx.MultiInput<framework::Tensor>("X");
+    auto out_var_names = ctx.Outputs(framework::GradVarName("X"));
     auto outs = ctx.MultiOutput<framework::Tensor>(framework::GradVarName("X"));
     int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
 
+    // get output tensor that the name is not kEmptyVarName
+    std::vector<framework::Tensor*> outputs;
+    for (size_t j = 0; j < outs.size(); ++j) {
+      if (out_var_names[j] != framework::kEmptyVarName) {
+        outs[j]->mutable_data<T>(ctx.GetPlace());
+        outputs.push_back(outs[j]);
+      } else {
+        outputs.push_back(nullptr);
+      }
+    }
+
     // Sometimes direct copies will be faster, this maybe need deeply analysis.
     if (axis == 0 && outs.size() < 10) {
       size_t input_offset = 0;
-      auto in_stride = framework::stride_numel(in->dims());
+      const auto in_stride = framework::stride_numel(out_grad->dims());
 
-      for (auto& out : outs) {
-        out->mutable_data<T>(ctx.GetPlace());
-        auto out_stride = framework::stride_numel(out->dims());
-        StridedNumelCopyWithAxis<T>(ctx.device_context(), axis, out->data<T>(),
-                                    out_stride, in->data<T>() + input_offset,
-                                    in_stride, out_stride[axis]);
+      for (size_t i = 0; i < outs.size(); ++i) {
+        auto out_stride = framework::stride_numel(ins[i]->dims());
+        auto* out = outputs[i];
+        if (out != nullptr) {
+          StridedNumelCopyWithAxis<T>(
+              ctx.device_context(), axis, out->data<T>(), out_stride,
+              out_grad->data<T>() + input_offset, in_stride, out_stride[axis]);
+        }
         input_offset += out_stride[axis];
       }
     } else {
-      std::vector<framework::Tensor> outputs(outs.size());
-      for (size_t j = 0; j < outs.size(); ++j) {
-        outs[j]->mutable_data<T>(ctx.GetPlace());
-        outputs[j] = *outs[j];
-      }
-
       auto& dev_ctx = ctx.template device_context<DeviceContext>();
       paddle::operators::math::ConcatGradFunctor<DeviceContext, T>
           concat_grad_functor;
-      concat_grad_functor(dev_ctx, *in, static_cast<int>(axis), &outputs);
+      concat_grad_functor(dev_ctx, *out_grad, ins, static_cast<int>(axis),
+                          &outputs);
     }
   }
 };

paddle/fluid/operators/detection_map_op.cc

@@ -175,12 +175,12 @@ class DetectionMAPOpMaker : public framework::OpProtoAndCheckerMaker {
     AddComment(R"DOC(
 Detection mAP evaluate operator.
 The general steps are as follows. First, calculate the true positive and
- false positive according to the input of detection and labels, then
- calculate the mAP evaluate value.
- Supporting '11 point' and 'integral' mAP algorithm. Please get more information
- from the following articles:
- https://sanchom.wordpress.com/tag/average-precision/
- https://arxiv.org/abs/1512.02325
+false positive according to the input of detection and labels, then
+calculate the mAP evaluate value.
+Supporting '11 point' and 'integral' mAP algorithm. Please get more information
+from the following articles:
+https://sanchom.wordpress.com/tag/average-precision/
+https://arxiv.org/abs/1512.02325
 
 )DOC");
   }

paddle/fluid/operators/gaussian_random_mkldnn_op.cc

+/* 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. */
+
+#include <string>
+#include "paddle/fluid/operators/mean_op.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::DataLayout;
+template <typename T>
+class GaussianMKLDNNKernel : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    float mean = context.Attr<float>("mean");
+    float std = context.Attr<float>("std");
+    auto* tensor = context.Output<framework::Tensor>("Out");
+    T* data = tensor->mutable_data<T>(context.GetPlace());
+
+    unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
+    std::minstd_rand engine;
+    if (seed == 0) {
+      seed = std::random_device()();
+    }
+    engine.seed(seed);
+    std::normal_distribution<T> dist(mean, std);
+    int64_t size = tensor->numel();
+    for (int64_t i = 0; i < size; ++i) {
+      data[i] = dist(engine);
+    }
+
+    // The format of output is set as the mkldnn's format
+    // TODO(@mozga-intel) The format of matrix sets inside the another layers.
+    tensor->set_layout(DataLayout::kMKLDNN);
+    tensor->set_format(mkldnn::memory::format::oihw);
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_KERNEL(gaussian_random, MKLDNN, ::paddle::platform::CPUPlace,
+                   ops::GaussianMKLDNNKernel<float>);

paddle/fluid/operators/gaussian_random_op.cc

@@ -15,6 +15,10 @@ limitations under the License. */
 #include <random>
 #include "paddle/fluid/framework/op_registry.h"
 
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
+
 namespace paddle {
 namespace operators {
 
@@ -62,9 +66,20 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
  protected:
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
+    framework::LibraryType library{framework::LibraryType::kPlain};
+    framework::DataLayout layout{framework::DataLayout::kAnyLayout};
+
+#ifdef PADDLE_WITH_MKLDNN
+    if (library == framework::LibraryType::kPlain &&
+        platform::CanMKLDNNBeUsed(ctx)) {
+      library = framework::LibraryType::kMKLDNN;
+      layout = framework::DataLayout::kMKLDNN;
+    }
+#endif
+
     return framework::OpKernelType(
         static_cast<framework::proto::VarType::Type>(ctx.Attr<int>("dtype")),
-        ctx.device_context());
+        ctx.device_context(), layout, library);
   }
 };
 
@@ -95,7 +110,9 @@ class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker {
                  "(int, default 5(FP32)) "
                  "Output data type.")
         .SetDefault(framework::proto::VarType::FP32);
-
+    AddAttr<bool>("use_mkldnn",
+                  "(bool, default false) Only used in mkldnn kernel")
+        .SetDefault(false);
     AddComment(R"DOC(
 GaussianRandom Operator.
 

paddle/fluid/operators/math/concat.cc

@@ -70,35 +70,40 @@ template <typename T>
 class ConcatGradFunctor<platform::CPUDeviceContext, T> {
  public:
   void operator()(const platform::CPUDeviceContext& context,
-                  const framework::Tensor& input, const int axis,
-                  std::vector<framework::Tensor>* outputs) {
+                  const framework::Tensor& input,
+                  const std::vector<const framework::Tensor*>& ref_inputs,
+                  const int axis, std::vector<framework::Tensor*>* outputs) {
     // TODO(zcd): Add input data validity checking
-    int num = outputs->size();
+    size_t num = outputs->size();
 
     int input_rows = 1;
-    auto dim_0 = outputs->at(0).dims();
+    auto dim_0 = ref_inputs[0]->dims();
     for (int i = 0; i < axis; ++i) {
       input_rows *= dim_0[i];
     }
+
     int input_cols = 0;
 
     std::vector<int64_t> output_cols(outputs->size());
-    for (int i = 0; i < num; ++i) {
-      int t_cols = outputs->at(i).numel() / input_rows;
+    for (size_t i = 0; i < num; ++i) {
+      int t_cols = ref_inputs[i]->numel() / input_rows;
       input_cols += t_cols;
       output_cols[i] = t_cols;
     }
     auto cpu_place = boost::get<platform::CPUPlace>(context.GetPlace());
 
     // computation
-    for (int k = 0; k < input_rows; ++k) {
+    for (size_t k = 0; k < input_rows; ++k) {
       const T* src_ptr = input.data<T>() + k * input_cols;
       int col_idx = 0;
       for (int j = 0; j < num; ++j) {
         int col_len = output_cols[j];
-        T* dst_ptr = outputs->at(j).data<T>() + k * col_len;
+        auto* out_tensor = outputs->at(j);
+        if (out_tensor != nullptr) {
+          T* dst_ptr = out_tensor->data<T>() + k * col_len;
           memory::Copy(cpu_place, dst_ptr, cpu_place, src_ptr + col_idx,
                        sizeof(T) * col_len);
+        }
         col_idx += col_len;
       }
     }

paddle/fluid/operators/math/concat.cu

@@ -22,43 +22,24 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template <typename T>
-__device__ T upper_bound(const T* first, T count, T val) {
-  const T* orig = first;
-  const T* it = nullptr;
-  T step = 0;
-  while (count > 0) {
-    it = first;
-    step = count / 2;
-    it += step;
-    if (!(val < *it)) {
-      first = ++it;
-      count -= step + 1;
-    } else {
-      count = step;
-    }
-  }
-  return first - orig;
-}
-
 template <typename T>
 __global__ void KernelConcat(T** inputs, const int* input_cols, int col_size,
                              const int output_rows, const int output_cols,
                              T* output) {
   int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
-  int segment = upper_bound<int>(input_cols, col_size, tid_x) - 1;
-
-  int curr_offset = input_cols[segment];
-  int curr_segment = segment;
+  int curr_segment = 0;
+  int curr_offset = input_cols[0];
   for (; tid_x < output_cols; tid_x += blockDim.x * gridDim.x) {
-    T curr_col_offset;
-    while ((curr_col_offset = input_cols[curr_segment + 1]) <= tid_x) {
+    int curr_col_offset = input_cols[curr_segment + 1];
+    while (curr_col_offset <= tid_x) {
       curr_offset = curr_col_offset;
       ++curr_segment;
+      curr_col_offset = input_cols[curr_segment + 1];
     }
 
     int local_col = tid_x - curr_offset;
     int segment_width = curr_col_offset - curr_offset;
+
     T* input_ptr = inputs[curr_segment];
     int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
     for (; tid_y < output_rows; tid_y += blockDim.y * gridDim.y)
@@ -89,24 +70,26 @@ __global__ void KernelConcatGrad(const T* input_data, const int in_row,
                                  const int in_col, const int* out_cols,
                                  int out_cols_size, T** outputs_data) {
   int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
-  int segment = upper_bound<int>(out_cols, out_cols_size, tid_x) - 1;
-  int curr_offset = out_cols[segment];
-  int curr_segment = segment;
+  int curr_segment = 0;
+  int curr_offset = out_cols[0];
   for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) {
-    T curr_col_offset;
-    while ((curr_col_offset = out_cols[curr_segment + 1]) <= tid_x) {
+    int curr_col_offset = out_cols[curr_segment + 1];
+    while (curr_col_offset <= tid_x) {
       curr_offset = curr_col_offset;
       ++curr_segment;
+      curr_col_offset = out_cols[curr_segment + 1];
     }
 
     int local_col = tid_x - curr_offset;
     int segment_width = curr_col_offset - curr_offset;
     T* output_ptr = outputs_data[curr_segment];
+    if (output_ptr != nullptr) {
       int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
       for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y)
         output_ptr[tid_y * segment_width + local_col] =
             input_data[tid_y * in_col + tid_x];
     }
+  }
 }
 
 template <typename T>
@@ -118,11 +101,13 @@ __global__ void KernelConcatGrad(const T* input_data, const int in_row,
     int split = tid_x / fixed_out_col;
     int in_offset = tid_x - split * fixed_out_col;
     T* output_ptr = outputs_data[split];
+    if (output_ptr != nullptr) {
       int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
       for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y)
         output_ptr[tid_y * fixed_out_col + in_offset] =
             input_data[tid_y * in_col + tid_x];
     }
+  }
 }
 
 /*
@@ -203,17 +188,18 @@ template <typename T>
 class ConcatGradFunctor<platform::CUDADeviceContext, T> {
  public:
   void operator()(const platform::CUDADeviceContext& context,
-                  const framework::Tensor& input, const int axis,
-                  std::vector<framework::Tensor>* outputs) {
+                  const framework::Tensor& input,
+                  const std::vector<const framework::Tensor*>& ref_inputs,
+                  const int axis, std::vector<framework::Tensor*>* outputs) {
     // TODO(zcd): Add input data validity checking
     int o_num = outputs->size();
     int out_row = 1;
-    auto dim_0 = outputs->at(0).dims();
+    auto dim_0 = ref_inputs[0]->dims();
     for (int i = 0; i < axis; ++i) {
       out_row *= dim_0[i];
     }
 
-    int out_col = outputs->at(0).numel() / out_row;
+    int out0_col = ref_inputs[0]->numel() / out_row;
     int in_col = 0, in_row = out_row;
     bool sameShape = true;
 
@@ -223,13 +209,17 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
 
     outputs_cols[0] = 0;
     for (int i = 0; i < o_num; ++i) {
-      int t_col = outputs->at(i).numel() / out_row;
+      int t_col = ref_inputs.at(i)->numel() / out_row;
       if (sameShape) {
-        if (t_col != out_col) sameShape = false;
+        if (t_col != out0_col) sameShape = false;
       }
       in_col += t_col;
       outputs_cols[i + 1] = in_col;
-      outputs_ptr[i] = outputs->at(i).data<T>();
+      if (outputs->at(i) != nullptr) {
+        outputs_ptr[i] = outputs->at(i)->data<T>();
+      } else {
+        outputs_ptr[i] = nullptr;
+      }
     }
 
     T** dev_out_gpu_data =
@@ -255,7 +245,7 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
 
     if (sameShape) {
       KernelConcatGrad<<<grid_size, block_size, 0, context.stream()>>>(
-          input.data<T>(), in_row, in_col, out_col, dev_out_gpu_data);
+          input.data<T>(), in_row, in_col, out0_col, dev_out_gpu_data);
     } else {
       const int* dev_outs_col_data = outputs_cols.CUDAData(context.GetPlace());
       KernelConcatGrad<<<grid_size, block_size, 0, context.stream()>>>(

paddle/fluid/operators/math/concat.h

@@ -57,7 +57,8 @@ template <typename DeviceContext, typename T>
 class ConcatGradFunctor {
  public:
   void operator()(const DeviceContext& context, const framework::Tensor& input,
-                  const int axis, std::vector<framework::Tensor>* outputs);
+                  const std::vector<const framework::Tensor*>& ref_inputs,
+                  const int axis, std::vector<framework::Tensor*>* outputs);
 };
 
 }  // namespace math

paddle/fluid/operators/parallel_do_op.cc

@@ -295,7 +295,7 @@ class ParallelDoGradOp : public framework::OperatorBase {
 
         auto sum_op = framework::OpRegistry::CreateOp(
             "sum", {{"X", {s, tmp_name}}}, {{"Out", {s}}},
-            framework::AttributeMap{});
+            framework::AttributeMap{{"use_mkldnn", {false}}});
         VLOG(10) << sum_op->DebugStringEx(sub_scopes[0]);
         sum_op->Run(*sub_scopes[0], places[0]);
         WaitOnPlace(places[0]);

paddle/fluid/operators/recurrent_op.cc

@@ -429,7 +429,8 @@ class RecurrentGradOp : public RecurrentBase {
 
           auto sum_op = framework::OpRegistry::CreateOp(
               "sum", {{"X", {pg_names[param_id], new_inside_name}}},
-              {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{});
+              {{"Out", {pg_names[param_id]}}},
+              framework::AttributeMap{{"use_mkldnn", {false}}});
           sum_op->Run(cur_scope, place);
 
           cur_scope.Rename(new_inside_name, inside_grad_name);

paddle/fluid/operators/sum_mkldnn_op.cc

+//   Copyright (c) 2018 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.
+
+/*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 "mkldnn.hpp"
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/operators/math/selected_rows_functor.h"
+#include "paddle/fluid/operators/sum_op.h"
+#include "paddle/fluid/platform/device_context.h"
+#include "paddle/fluid/platform/mkldnn_helper.h"
+
+namespace paddle {
+namespace operators {
+
+using paddle::framework::Tensor;
+using paddle::platform::MKLDNNDeviceContext;
+using paddle::platform::CPUDeviceContext;
+using framework::DataLayout;
+using mkldnn::memory;
+using mkldnn::primitive;
+using mkldnn::stream;
+using mkldnn::sum;
+using mkldnn::reorder;
+using platform::to_void_cast;
+
+template <typename T>
+class SumMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
+                   "It must use CPUPlace.");
+    auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
+    const auto& mkldnn_engine = dev_ctx.GetEngine();
+    auto in_vars = ctx.MultiInputVar("X");
+
+    const int N = in_vars.size();
+    auto out_var = ctx.OutputVar("Out");
+    bool in_place = out_var == in_vars[0];
+
+    if (out_var->IsType<framework::LoDTensor>()) {
+      LoDTensor* output = ctx.Output<LoDTensor>("Out");
+      T* output_data = output->mutable_data<T>(ctx.GetPlace());
+
+      std::vector<int> dst_tz = framework::vectorize2int(output->dims());
+      auto src_tz = dst_tz;
+      memory::format output_format{memory::format::format_undef};
+      std::vector<float> scales;
+      std::vector<memory::primitive_desc> srcs_mpd;
+      std::vector<mkldnn::memory> srcs_mem;
+
+      PADDLE_ENFORCE(in_vars[0]->IsType<LoDTensor>(),
+                     "Input[0] must be LoDTensors");
+      auto& input0 = in_vars[0]->Get<LoDTensor>();
+      PADDLE_ENFORCE(input0.layout() == DataLayout::kMKLDNN &&
+                         input0.format() != memory::format::format_undef,
+                     "Wrong layout/format for inputs[0]");
+
+      memory::format input_format = input0.format();
+
+      if (src_tz.size() == 1 && (input_format == memory::format::nchw ||
+                                 input_format == memory::format::nhwc)) {
+        input_format = memory::format::x;
+      }
+      if (src_tz.size() == 2 && (input_format == memory::format::nchw ||
+                                 input_format == memory::format::nhwc)) {
+        input_format = memory::format::nc;
+      }
+
+      for (int i = in_place ? 1 : 0; i < N; i++) {
+        PADDLE_ENFORCE(in_vars[i]->IsType<LoDTensor>(),
+                       "all inputs must be all LoDTensors");
+        auto& input = in_vars[i]->Get<LoDTensor>();
+        PADDLE_ENFORCE(input.layout() == DataLayout::kMKLDNN &&
+                           input.format() != memory::format::format_undef,
+                       "Wrong layout/format for inputs");
+
+        if (input.numel() == 0) {
+          continue;
+        }
+
+        const T* input_data = input.data<T>();
+
+        auto src_md =
+            memory::desc(src_tz, memory::data_type::f32, input_format);
+        auto src_mpd = memory::primitive_desc(src_md, mkldnn_engine);
+        auto src_mem = memory(src_mpd, to_void_cast(input_data));
+        srcs_mpd.push_back(src_mpd);
+        srcs_mem.push_back(src_mem);
+        scales.push_back(1.0);
+      }
+
+      auto dst_md =
+          memory::desc(dst_tz, memory::data_type::f32, memory::format::any);
+
+      auto sum_pd = sum::primitive_desc(dst_md, scales, srcs_mpd);
+
+      std::shared_ptr<memory> dst_mem;
+      if (in_place) {
+        dst_mem.reset(new memory(sum_pd.dst_primitive_desc()));
+      } else {
+        dst_mem.reset(new memory(sum_pd.dst_primitive_desc(), output_data));
+      }
+      std::vector<mkldnn::primitive::at> inputs;
+      for (size_t i = 0; i < srcs_mem.size(); ++i) {
+        inputs.push_back(srcs_mem[i]);
+      }
+
+      auto sum_prim = mkldnn::sum(sum_pd, inputs, *dst_mem);
+      output_format = (memory::format)platform::GetMKLDNNFormat(sum_pd);
+
+      primitive reorder_prim;
+      std::shared_ptr<memory> target_mem;
+      if (in_place) {
+        output_format = input_format;
+        target_mem.reset(new memory(
+            {{{src_tz}, memory::data_type::f32, output_format}, mkldnn_engine},
+            output_data));
+        reorder_prim = reorder(*dst_mem, *target_mem);
+      }
+
+      std::vector<primitive> pipeline;
+      pipeline.push_back(sum_prim);
+      if (in_place) pipeline.push_back(reorder_prim);
+      stream(stream::kind::eager).submit(pipeline).wait();
+
+      output->set_layout(DataLayout::kMKLDNN);
+      output->set_format(output_format);
+    } else if (out_var->IsType<framework::SelectedRows>()) {
+      // TODO(@mozga-intel) Add MKLDNN SelectedRows support
+      std::unique_ptr<framework::SelectedRows> in0;
+      if (in_place) {
+        // If is in_place, we store the input[0] to in0
+        auto& in_sel0 = in_vars[0]->Get<SelectedRows>();
+        auto& rows = in_sel0.rows();
+        in0.reset(new framework::SelectedRows(rows, in_sel0.height()));
+        in0->mutable_value()->ShareDataWith(in_sel0.value());
+      }
+
+      auto get_selected_row = [&](size_t i) -> const SelectedRows& {
+        if (i == 0 && in0) {
+          return *in0.get();
+        } else {
+          return in_vars[i]->Get<SelectedRows>();
+        }
+      };
+      auto* out = ctx.Output<SelectedRows>("Out");
+      out->mutable_rows()->clear();
+      auto* out_value = out->mutable_value();
+
+      // Runtime InferShape
+      size_t first_dim = 0;
+      for (int i = 0; i < N; i++) {
+        auto& sel_row = get_selected_row(i);
+        first_dim += sel_row.rows().size();
+      }
+      auto in_dim =
+          framework::vectorize(get_selected_row(N - 1).value().dims());
+      in_dim[0] = static_cast<int64_t>(first_dim);
+
+      out_value->Resize(framework::make_ddim(in_dim));
+
+      // if all the input sparse vars are empty, no need to
+      // merge these vars.
+      if (first_dim == 0UL) {
+        return;
+      }
+      out_value->mutable_data<T>(ctx.GetPlace());
+      math::SelectedRowsAddTo<CPUDeviceContext, T> functor;
+      int64_t offset = 0;
+      for (int i = 0; i < N; i++) {
+        auto& sel_row = get_selected_row(i);
+        if (sel_row.rows().size() == 0) {
+          continue;
+        }
+        PADDLE_ENFORCE_EQ(out->height(), sel_row.height());
+        functor(ctx.template device_context<CPUDeviceContext>(), sel_row,
+                offset, out);
+        offset += sel_row.value().numel();
+      }
+    } else if (out_var->IsType<framework::LoDTensorArray>()) {
+      // TODO(@mozga-intel) Add MKLDNN LoDTensorArray support
+      auto& out_array = *out_var->GetMutable<framework::LoDTensorArray>();
+      for (size_t i = in_place ? 1 : 0; i < in_vars.size(); ++i) {
+        PADDLE_ENFORCE(in_vars[i]->IsType<framework::LoDTensorArray>(),
+                       "Only support all inputs are TensorArray");
+        auto& in_array = in_vars[i]->Get<framework::LoDTensorArray>();
+
+        for (size_t i = 0; i < in_array.size(); ++i) {
+          if (in_array[i].numel() != 0) {
+            if (i >= out_array.size()) {
+              out_array.resize(i + 1);
+            }
+            if (out_array[i].numel() == 0) {
+              framework::TensorCopy(in_array[i], in_array[i].place(),
+                                    ctx.device_context(), &out_array[i]);
+              out_array[i].set_lod(in_array[i].lod());
+            } else {
+              PADDLE_ENFORCE(out_array[i].lod() == in_array[i].lod());
+              auto in = EigenVector<T>::Flatten(in_array[i]);
+              auto result = EigenVector<T>::Flatten(out_array[i]);
+              result.device(*ctx.template device_context<MKLDNNDeviceContext>()
+                                 .eigen_device()) = result + in;
+            }
+          }
+        }
+      }
+    } else {
+      PADDLE_THROW("Unexpected branch, output variable type is %s",
+                   out_var->Type().name());
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_KERNEL(sum, MKLDNN, ::paddle::platform::CPUPlace,
+                   paddle::operators::SumMKLDNNOpKernel<float>);

paddle/fluid/operators/sum_op.cc

@@ -18,6 +18,10 @@ limitations under the License. */
 #include "paddle/fluid/framework/var_type_inference.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
 
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
+
 namespace paddle {
 namespace operators {
 using framework::Tensor;
@@ -63,6 +67,18 @@ class SumOp : public framework::OperatorWithKernel {
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     auto x_vars = ctx.MultiInputVar("X");
+
+    framework::LibraryType library{framework::LibraryType::kPlain};
+    framework::DataLayout layout{framework::DataLayout::kAnyLayout};
+
+#ifdef PADDLE_WITH_MKLDNN
+    if (library == framework::LibraryType::kPlain &&
+        platform::CanMKLDNNBeUsed(ctx)) {
+      library = framework::LibraryType::kMKLDNN;
+      layout = framework::DataLayout::kMKLDNN;
+    }
+#endif
+
     if (x_vars[0]->IsType<framework::LoDTensor>()) {
       int dtype = -1;
       for (auto& x_var : x_vars) {
@@ -80,26 +96,27 @@ class SumOp : public framework::OperatorWithKernel {
                         "Sum operator should have at least one tensor");
 
       return framework::OpKernelType(
-          static_cast<framework::proto::VarType::Type>(dtype),
-          ctx.device_context());
+          static_cast<framework::proto::VarType::Type>(dtype), ctx.GetPlace(),
+          layout, library);
     } else if (x_vars[0]->IsType<framework::SelectedRows>()) {
       for (auto& var : x_vars) {
         auto& value = var->Get<framework::SelectedRows>().value();
         if (value.IsInitialized()) {
           return framework::OpKernelType(framework::ToDataType(value.type()),
-                                         ctx.device_context());
+                                         ctx.device_context(), layout, library);
         }
       }
       // if input sparse vars are not initialized, use an default kernel type.
       return framework::OpKernelType(framework::proto::VarType::FP32,
-                                     ctx.device_context());
+                                     ctx.device_context(), layout, library);
     } else if (x_vars[0]->IsType<framework::LoDTensorArray>()) {
       for (auto& x_var : x_vars) {
         auto& array = x_var->Get<framework::LoDTensorArray>();
         for (auto& each : array) {
           if (each.numel() != 0) {
             return framework::OpKernelType(framework::ToDataType(each.type()),
-                                           ctx.device_context());
+                                           ctx.device_context(), layout,
+                                           library);
           }
         }
       }
@@ -116,6 +133,9 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X", "(vector<Tensor>) The input tensors of sum operator.")
         .AsDuplicable();
     AddOutput("Out", "(Tensor) The output tensor of sum operator.").Reuse("X");
+    AddAttr<bool>("use_mkldnn",
+                  "(bool, default false) Only used in mkldnn kernel")
+        .SetDefault(false);
     AddComment(R"DOC(
 Sum operator.
 
@@ -132,7 +152,6 @@ class SumOpVarTypeInference : public framework::VarTypeInference {
                   framework::BlockDesc* block) const override {
     auto& inputs = op_desc.Input("X");
     auto var_type = framework::proto::VarType::SELECTED_ROWS;
-
     for (auto& name : op_desc.Input("X")) {
       VLOG(10) << name << " "
                << block->FindRecursiveOrCreateVar(name).GetType();
@@ -206,6 +225,7 @@ namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker,
                   ops::SumOpVarTypeInference);
+
 REGISTER_OP_CPU_KERNEL(
     sum, ops::SumKernel<paddle::platform::CPUDeviceContext, float>,
     ops::SumKernel<paddle::platform::CPUDeviceContext, double>,

paddle/fluid/operators/while_op.cc

@@ -203,11 +203,11 @@ class WhileGradOp : public framework::OperatorBase {
                 ->set_lod(inside_tensor.lod());
           }
         }
-
         auto new_inside_name = cur_scope.Rename(inside_grad_name);
         auto sum_op = framework::OpRegistry::CreateOp(
             "sum", {{"X", {pg_names[param_id], new_inside_name}}},
-            {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{});
+            {{"Out", {pg_names[param_id]}}},
+            framework::AttributeMap{{"use_mkldnn", {false}}});
         sum_op->Run(cur_scope, dev_place);
         cur_scope.Rename(new_inside_name, inside_grad_name);
       }

paddle/fluid/platform/cpu_info.cc

@@ -28,9 +28,15 @@ DEFINE_double(fraction_of_cpu_memory_to_use, 1,
               "Default use 100% of CPU memory for PaddlePaddle,"
               "reserve the rest for page tables, etc");
 
-DEFINE_uint64(
-    initial_cpu_memory_in_mb, 500,
-    "Default initial 500MB of CPU memory for PaddlePaddle, in MD unit.");
+DEFINE_uint64(initial_cpu_memory_in_mb,
+#ifdef PADDLE_WITH_MKLDNN
+              /* Aligned with mozga-intel, MKLDNN need at least 5000 MB
+               * to obtain the best performance*/
+              5000,
+#else
+              500,
+#endif
+              "Initial CPU memory for PaddlePaddle, in MD unit.");
 
 DEFINE_double(
     fraction_of_cuda_pinned_memory_to_use, 0.5,
@@ -59,10 +65,7 @@ inline size_t CpuTotalPhysicalMemory() {
 size_t CpuMaxAllocSize() {
   // For distributed systems, it requires configuring and limiting
   // the fraction of memory to use.
-  return std::min(
-      static_cast<size_t>(FLAGS_fraction_of_cpu_memory_to_use *
-                          CpuTotalPhysicalMemory()),
-      static_cast<size_t>(FLAGS_initial_cpu_memory_in_mb * 1 << 20));
+  return FLAGS_fraction_of_cpu_memory_to_use * CpuTotalPhysicalMemory();
 }
 
 size_t CpuMinChunkSize() {
@@ -71,8 +74,11 @@ size_t CpuMinChunkSize() {
 }
 
 size_t CpuMaxChunkSize() {
-  // Allow to allocate the maximum chunk size is roughly 3% of CPU memory.
-  return CpuMaxAllocSize() / 32;
+  // Allow to allocate the maximum chunk size is roughly 3% of CPU memory,
+  // or the initial_cpu_memory_in_mb.
+  return std::min(
+      static_cast<size_t>(CpuMaxAllocSize() / 32),
+      static_cast<size_t>(FLAGS_initial_cpu_memory_in_mb * 1 << 20));
 }
 
 size_t CUDAPinnedMaxAllocSize() {

paddle/fluid/platform/mkldnn_helper.h

@@ -99,5 +99,11 @@ inline mkldnn::memory::format GetMKLDNNFormat(const mkldnn::memory memory) {
       memory.get_primitive_desc().desc().data.format);
 }
 
+inline mkldnn::memory::format GetMKLDNNFormat(
+    const mkldnn::sum::primitive_desc& memory) {
+  return static_cast<mkldnn::memory::format>(
+      memory.dst_primitive_desc().desc().data.format);
+}
+
 }  // namespace platform
 }  // namespace paddle

paddle/testing/paddle_gtest_main.cc

@@ -30,8 +30,9 @@ int main(int argc, char** argv) {
   new_argv.push_back(
       strdup("--tryfromenv=fraction_of_gpu_memory_to_use,use_pinned_memory"));
 #else
-  new_argv.push_back(strdup("--tryfromenv=use_pinned_memory,use_mkldnn"));
-  new_argv.push_back(strdup("--undefok=use_mkldnn"));
+  new_argv.push_back(strdup(
+      "--tryfromenv=use_pinned_memory,use_mkldnn,initial_cpu_memory_in_mb"));
+  new_argv.push_back(strdup("--undefok=use_mkldnn,initial_cpu_memory_in_mb"));
 #endif
   int new_argc = static_cast<int>(new_argv.size());
   char** new_argv_address = new_argv.data();

python/paddle/fluid/__init__.py

@@ -117,7 +117,7 @@ def __bootstrap__():
 
     read_env_flags = [
         'use_pinned_memory', 'check_nan_inf', 'benchmark', 'warpctc_dir',
-        'eager_delete_scope', 'use_mkldnn'
+        'eager_delete_scope', 'use_mkldnn', 'initial_cpu_memory_in_mb'
     ]
     if core.is_compiled_with_cuda():
         read_env_flags += [

python/paddle/fluid/average.py

@@ -36,6 +36,25 @@ def _is_number_or_matrix_(var):
 
 
 class WeightedAverage(object):
+    """
+    Calculate weighted average.
+
+    The average calculating is accomplished via Python totally. 
+    They do not change Paddle's Program, nor do anything to
+    modify NN model's configuration. They are completely 
+    wrappers of Python functions.
+
+    Examples:
+        .. code-block:: python
+            avg = fluid.average.WeightedAverage()
+            avg.add(value=2.0, weight=1)
+            avg.add(value=4.0, weight=2)
+            avg.eval()
+
+            # The result is 3.333333333.
+            # For (2.0 * 1 + 4.0 * 2) / (1 + 2) = 3.333333333
+    """
+
     def __init__(self):
         warnings.warn(
             "The %s is deprecated, please use fluid.metrics.Accuracy instead." %

python/paddle/fluid/backward.py

@@ -132,9 +132,9 @@ def _addup_repetitive_outputs_(op_descs):
     for idx, op_desc in enumerate(op_descs):
         for var_name in op_desc.input_arg_names():
             if len(renamed_vars[var_name]) > 1:
-                pending_sum_ops.append(
-                    (_create_op_desc_("sum", {"X": renamed_vars[var_name]},
-                                      {"Out": [var_name]}, {}), idx))
+                pending_sum_ops.append((_create_op_desc_(
+                    "sum", {"X": renamed_vars[var_name]}, {"Out": [var_name]},
+                    {"use_mkldnn": False}), idx))
                 renamed_vars[var_name] = [var_name]
         for var_name in op_desc.output_arg_names():
             if var_name == core.empty_var_name(
@@ -161,8 +161,9 @@ def _addup_repetitive_outputs_(op_descs):
                 renamed_vars[var_name].append(new_name)
     for var_name, inputs in renamed_vars.iteritems():
         if len(inputs) > 1:
-            pending_sum_ops.append((_create_op_desc_(
-                "sum", {"X": inputs}, {"Out": [var_name]}, {}), len(op_descs)))
+            pending_sum_ops.append(
+                (_create_op_desc_("sum", {"X": inputs}, {"Out": [var_name]},
+                                  {"use_mkldnn": False}), len(op_descs)))
     # sum_op descs are sorted according to their insert position
     for p in reversed(pending_sum_ops):
         op_descs.insert(p[1], p[0])
@@ -434,18 +435,65 @@ def _get_stop_gradients_(program):
 def append_backward(loss, parameter_list=None, no_grad_set=None,
                     callbacks=None):
     """
-    Append backward part to main_program
+    Append backward part to main_program.
 
-    Args:
-        loss(Variable): The variable generated by cost function.
-        parameter_list(list[string]): Parameters that need to be updated by
-            optimizer. If None, it means all parameters need to be updated.
-        no_grad_set(set): Variables that have no gradients in Block 0.
-            All variables with `step_gradient=True` from all blocks will be
-            automatically added.
+    A complete neural network training is made up of forward and backward 
+    propagation. However, when we configure a network, we only need to 
+    specify its forwrd part. The backward part is generated automatically 
+    according to the forward part by this function.
 
-    Return:
-        (list[(Variable,Variable)]): list of (parameter, gradient) pair.
+    In most cases, users do not need to invoke this function manually. It 
+    will be automatically invoked by the optimizer's `minimize` function.
+
+    Args:
+        loss(Variable): The loss variable of the network.
+        parameter_list(list[string]|None): Names of parameters that need 
+                                           to be updated by optimizers. 
+                                           If it is None, all parameters 
+                                           will be updated.
+                                           Default: None
+        no_grad_set(set|None): Variables in the Block 0 whose gradients 
+                               should be ignored. All variables with 
+                               `step_gradient=True` from all blocks will 
+                               be automatically added into this set.
+                               Default: None
+        callbacks(list[callable object]|None): The callbacks are used for 
+                                               doing some custom jobs during 
+                                               backward part building. All 
+                                               callable objects in it will 
+                                               be invoked once each time a 
+                                               new gradient operator is added 
+                                               into the program. The callable 
+                                               object must has two input 
+                                               parameters: 'block' and 'context'. 
+                                               The 'block' is the block which 
+                                               the new gradient operator will 
+                                               be added to. The 'context' is a 
+                                               map, whose keys are gradient 
+                                               variable names and values are 
+                                               corresponding original variables.
+                                               In addition to this, the 'context' 
+                                               has another special key-value pair: 
+                                               the key is string '__current_op_desc__' 
+                                               and the value is the op_desc of the 
+                                               gradient operator who has just 
+                                               triggered the callable object. 
+
+    Returns:
+        list[(Variable,Variable)]: Pairs of parameter and its 
+        corresponding gradients. The key is the parameter and the 
+        value is gradient variable.
+
+    Raises:
+        AssertionError: If `loss` is not an instance of Variable.
+
+    Examples:
+        .. code-block:: python
+
+            # network configuration code
+            # ...
+            avg_loss = fluid.layers.mean(loss)
+            param_grad_list = fluid.backward.append_backward(loss=avg_loss)
     """
     assert isinstance(loss, framework.Variable)
 

python/paddle/fluid/clip.py

@@ -24,8 +24,6 @@ __all__ = [
     'GradientClipByValue',
     'GradientClipByNorm',
     'GradientClipByGlobalNorm',
-    'append_gradient_clip_ops',
-    'error_clip_callback',
 ]
 
 
@@ -38,6 +36,25 @@ class BaseErrorClipAttr(object):
 
 
 class ErrorClipByValue(BaseErrorClipAttr):
+    """
+    Clips tensor values to the range [min, max].
+
+    Given a tensor t, this operation clips its value to min and max inplace.
+
+    - Any values less than min are set to min.
+    - Any values greater than max are set to max.
+
+    Args:
+        max (float): The maximum value to clip by.
+        min (float, optional): The minimum value to clip by. if not set by user, \
+        will be set to -max by framework.
+
+    Examples:
+        .. code-block:: python
+
+            var = fluid.framework.Variable(..., error_clip=ErrorClipByValue(max=5.0), ...)
+    """
+
     def __init__(self, max, min=None):
         max = float(max)
         if min is None:
@@ -99,6 +116,31 @@ class NullGradientClipAttr(BaseGradientClipAttr):
 
 
 class GradientClipByValue(BaseGradientClipAttr):
+    """
+    Clips gradient values to the range [min, max].
+
+    Given a tensor t, this operation clips its value to min and max inplace.
+
+    - Any values less than min are set to min.
+    - Any values greater than max are set to max.
+
+    Args:
+        max (float): The maximum value to clip by.
+        min (float, optional): The minimum value to clip by. if not set by user, \
+        will be set to -max by framework.
+
+    Examples:
+        .. code-block:: python
+
+            w_param_attrs = ParamAttr(name=None,
+              initializer=UniformInitializer(low=-1.0, high=1.0, seed=0),
+              learning_rate=1.0,
+              regularizer=L1Decay(1.0),
+              trainable=True,
+              clip=GradientClipByValue(-1.0, 1.0))
+            y_predict = fluid.layers.fc(input=x, size=1, param_attr=w_param_attrs)
+    """
+
     def __init__(self, max, min=None):
         max = float(max)
         if min is None:
@@ -120,6 +162,37 @@ class GradientClipByValue(BaseGradientClipAttr):
 
 
 class GradientClipByNorm(BaseGradientClipAttr):
+    """
+    Clips tensor values to a maximum L2-norm.
+
+    This operator limits the L2 norm of the input :math:`X` within :math:`max\_norm`.
+    If the L2 norm of :math:`X` is less than or equal to :math:`max\_norm`, :math:`Out`
+    will be the same as :math:`X`. If the L2 norm of :math:`X` is greater than
+    :math:`max\_norm`, :math:`X` will be linearly scaled to make the L2 norm of
+    :math:`Out` equal to :math:`max\_norm`, as shown in the following formula:
+
+    .. math::
+
+        Out = \\frac{max\_norm * X}{norm(X)},
+
+    where :math:`norm(X)` represents the L2 norm of :math:`X`.
+
+    Args:
+        clip_norm (float): The maximum norm value
+
+    Examples:
+        .. code-block:: python
+
+            w_param_attrs = ParamAttr(name=None,
+              initializer=UniformInitializer(low=-1.0, high=1.0, seed=0),
+              learning_rate=1.0,
+              regularizer=L1Decay(1.0),
+              trainable=True,
+              clip=GradientClipByNorm(clip_norm=2.0))
+            y_predict = fluid.layers.fc(input=x, size=1, param_attr=w_param_attrs)
+
+    """
+
     def __init__(self, clip_norm):
         self.clip_norm = clip_norm
 
@@ -135,6 +208,44 @@ class GradientClipByNorm(BaseGradientClipAttr):
 
 
 class GradientClipByGlobalNorm(BaseGradientClipAttr):
+    """
+    Clips values of multiple tensors by the ratio of the sum of their norms.
+
+    Given a list of tensors t_list, and a clipping ratio clip_norm, this
+    operation returns a list of clipped tensors list_clipped and the global
+    norm (global_norm) of all tensors in t_list.
+
+    To perform the clipping, the values :math:`t\_list[i]` are set to:
+
+    .. math::
+
+        t\_list[i] = t\_list[i] * \\frac{clip\_norm}{\max(global\_norm, clip\_norm)}
+
+    where:
+
+    .. math::
+
+        global\_norm = \sqrt{\sum_{i=0}^{N-1}(l2norm(t\_list[i]))^2}
+
+    If :math:`clip\_norm > global\_norm` then the entries in t_list remain as they are,
+    otherwise they're all shrunk by the global ratio.
+
+    Args:
+        clip_norm (float): The maximum norm value
+        group_name (str, optional): The group name for this clip.
+
+    Examples:
+        .. code-block:: python
+
+            p_g_clip = fluid.backward.append_backward(loss=avg_cost_clip)
+
+            with fluid.program_guard(main_program=prog_clip):
+                fluid.clip.set_gradient_clip(
+                    fluid.clip.GradientClipByGlobalNorm(clip_norm=2.0))
+                p_g_clip = fluid.clip.append_gradient_clip_ops(p_g_clip)
+
+    """
+
     def __init__(self, clip_norm, group_name="default_group"):
         if not isinstance(group_name, basestring):
             raise TypeError("'group_name' must be a basestring.")
@@ -184,13 +295,14 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr):
 def set_gradient_clip(clip, param_list=None, program=None):
     """
     To specify parameters that require gradient clip.
+
     Args:
         clip(BaseGradientClipAttr): An instance of some derived class of BaseGradientClipAttr,
                 which describes the type and detailed attributes of required gradient clip.
-            param_list(list, None by default): Parameters that require gradient clip. 
+        param_list(list(Variable)): Parameters that require gradient clip.
                 It can be a list of parameter or a list of parameter's name.
                 When it's None, all parameters in the program will be included.
-            program(Program, None by default): The program where parameters are. 
+        program(Program): The program where parameters are.
                 Will be the default main program when assigned with None.
     """
     if not isinstance(clip, BaseGradientClipAttr):

python/paddle/fluid/data_feeder.py

@@ -29,6 +29,13 @@ class DataToLoDTensorConverter(object):
         self.place = place
         self.lod_level = lod_level
         self.shape = shape
+        negtive_count = 0
+        for s in self.shape:
+            if s < 0:
+                negtive_count += 1
+            if negtive_count > 1:
+                self.shape = None
+                break
         if dtype == core.VarDesc.VarType.FP32:
             self.dtype = 'float32'
         elif dtype == core.VarDesc.VarType.INT64:
@@ -61,7 +68,9 @@ class DataToLoDTensorConverter(object):
                 self._feed_impl_(each_data, lod[1:], lod_level - 1)
 
     def done(self):
-        arr = numpy.array(self.data, dtype=self.dtype).reshape(self.shape)
+        arr = numpy.array(self.data, dtype=self.dtype)
+        if self.shape:
+            arr = arr.reshape(self.shape)
         t = core.LoDTensor()
         t.set(arr, self.place)
         if self.lod_level > 0:

python/paddle/fluid/framework.py

@@ -645,6 +645,12 @@ class Operator(object):
 
     def set_attr(self, name, val):
         self.attrs[name] = val
+        if isinstance(val, Block):
+            self.desc.set_block_attr(name, val.desc)
+        elif isinstance(val, core.BlockDesc) or \
+                isinstance(val, core.ProgramDesc):
+            self.desc.set_serialized_attr(name, val.serialize_to_string())
+        else:
             self.desc.set_attr(name, val)
 
     @property

python/paddle/fluid/inferencer.py

@@ -27,13 +27,30 @@ __all__ = ['Inferencer', ]
 
 
 class Inferencer(object):
-    def __init__(self, infer_func, param_path, place=None, parallel=False):
     """
-        :param infer_func: a function that will return predict Variable
-        :param param_path: the path where the inference model is saved by fluid.io.save_params
-        :param place: place to do the inference
-        :param parallel: use parallel_executor to run the inference, it will use multi CPU/GPU.
+    Inferencer High Level API.
+
+    Args:
+        infer_func (Python func): Infer function that will return predict Variable
+        param_path (str): The path where the inference model is saved by fluid.io.save_params
+        place (Place): place to do the inference
+        parallel (bool): use parallel_executor to run the inference, it will use multi CPU/GPU.
+
+    Examples:
+        .. code-block:: python
+
+            def inference_program():
+                x = fluid.layers.data(name='x', shape=[13], dtype='float32')
+                y_predict = fluid.layers.fc(input=x, size=1, act=None)
+                return y_predict
+
+            place = fluid.CPUPlace()
+            inferencer = fluid.Inferencer(
+                infer_func=inference_program, param_path="/tmp/model", place=place)
+
     """
+
+    def __init__(self, infer_func, param_path, place=None, parallel=False):
         self.param_path = param_path
         self.scope = core.Scope()
         self.parallel = parallel
@@ -60,9 +77,20 @@ class Inferencer(object):
 
     def infer(self, inputs, return_numpy=True):
         """
-        :param inputs: a map of {"input_name": input_var} that will be feed into the inference program
-        to get the predict value
-        :return: the predict value of the inference model
+        Do Inference for Inputs
+
+        Args:
+            inputs (map): a map of {"input_name": input_var} that will be feed into the inference program
+            return_numpy (bool): transform return value into numpy or not
+
+        Returns:
+            Tensor or Numpy: the predict value of the inference model for the inputs
+
+        Examples:
+            .. code-block:: python
+
+                tensor_x = numpy.random.uniform(0, 10, [batch_size, 13]).astype("float32")
+                results = inferencer.infer({'x': tensor_x})
         """
         if not isinstance(inputs, dict):
             raise ValueError(

python/paddle/fluid/initializer.py

@@ -19,26 +19,39 @@ from framework import convert_np_dtype_to_dtype_
 from core import VarDesc
 
 __all__ = [
-    'Constant', 'Uniform', 'Normal', 'Xavier', 'Bilinear', 'force_init_on_cpu',
-    'init_on_cpu', 'ConstantInitializer', 'UniformInitializer',
-    'NormalInitializer', 'XavierInitializer', 'BilinearInitializer'
+    'Constant', 'Uniform', 'Normal', 'Xavier', 'Bilinear', 'MSRA',
+    'force_init_on_cpu', 'init_on_cpu', 'ConstantInitializer',
+    'UniformInitializer', 'NormalInitializer', 'XavierInitializer',
+    'BilinearInitializer', 'MSRAInitializer'
 ]
 
 _force_init_on_cpu_ = False
 
 
 def force_init_on_cpu():
+    """
+    The flag of whether force to init variables on CPU.
+
+    Examples:
+        .. code-block:: python
+
+            if force_init_on_cpu():
+                pass
+
+    """
     return _force_init_on_cpu_
 
 
 @contextlib.contextmanager
 def init_on_cpu():
     """
-    Switch program with `with` statement
+    Force the variable to be inited on CPU.
 
     Examples:
-        >>> with init_on_cpu():
-        >>>   step = layers.create_global_var()
+        .. code-block:: python
+
+            with init_on_cpu():
+                step = layers.create_global_var()
 
     """
     global _force_init_on_cpu_
@@ -104,14 +117,18 @@ class Initializer(object):
 
 class ConstantInitializer(Initializer):
     """Implements the constant initializer
-    """
-
-    def __init__(self, value=0.0, force_cpu=False):
-        """Constructor for ConstantInitializer
 
     Args:
-            value: constant value to initialize the variable
+        value (float): constant value to initialize the variable
+
+    Examples:
+        .. code-block:: python
+
+            fc = fluid.layers.fc(input=x, size=10,
+                param_attr=fluid.initializer.Constant(value=2.0))
     """
+
+    def __init__(self, value=0.0, force_cpu=False):
         assert value is not None
         super(ConstantInitializer, self).__init__()
         self._value = value
@@ -146,16 +163,20 @@ class ConstantInitializer(Initializer):
 
 class UniformInitializer(Initializer):
     """Implements the random uniform distribution initializer
-    """
-
-    def __init__(self, low=-1.0, high=1.0, seed=0):
-        """Constructor for UniformInitializer
 
     Args:
-            low: lower boundary of the uniform distribution
-            high: upper boundary of the uniform distribution
-            seed: random seed
+        low (float): lower boundary of the uniform distribution
+        high (float): upper boundary of the uniform distribution
+        seed (int): random seed
+
+    Examples:
+        .. code-block:: python
+
+            fc = fluid.layers.fc(input=x, size=10,
+                param_attr=fluid.initializer.Uniform(low=-0.5, high=0.5))
     """
+
+    def __init__(self, low=-1.0, high=1.0, seed=0):
         assert low is not None
         assert high is not None
         assert high >= low
@@ -196,17 +217,21 @@ class UniformInitializer(Initializer):
 
 
 class NormalInitializer(Initializer):
-    """Implements the  random Normal(Gaussian) distribution initializer
-    """
-
-    def __init__(self, loc=0.0, scale=1.0, seed=0):
-        """Constructor for NormalInitializer
+    """Implements the Random Normal(Gaussian) distribution initializer
 
     Args:
-            loc: mean of the normal distribution
-            scale: standard deviation of the normal distribution
-            seed: random seed
+        loc (float): mean of the normal distribution
+        scale (float): standard deviation of the normal distribution
+        seed (int): random seed
+
+    Examples:
+        .. code-block:: python
+
+            fc = fluid.layers.fc(input=x, size=10,
+                param_attr=fluid.initializer.Normal(loc=0.0, scale=2.0))
     """
+
+    def __init__(self, loc=0.0, scale=1.0, seed=0):
         assert loc is not None
         assert scale is not None
         assert seed is not None
@@ -246,39 +271,49 @@ class NormalInitializer(Initializer):
 
 
 class XavierInitializer(Initializer):
-    """Implements the Xavier initializer
-
+    """
     This class implements the Xavier weight initializer from the paper
-    Understanding the difficulty of training deep feedforward neural
-    networks[1] by Xavier Glorot and Yoshua Bengio.
+    `Understanding the difficulty of training deep feedforward neural
+    networks <http://proceedings.mlr.press/v9/glorot10a/glorot10a.pdf>`_
+    by Xavier Glorot and Yoshua Bengio.
 
     This initializer is designed to keep the scale of the gradients
     approximately same in all the layers. In case of Uniform distribution,
-    the range is [-x, x], where x = sqrt(6 / (fan_in + fan_out)).
+    the range is [-x, x], where
+
+    .. math::
+
+        x = \sqrt{\\frac{6.0}{fan\_in + fan\_out}}
+
     In case of Normal distribution, the mean is 0 and the standard deviation
-    is sqrt(2/ (fan_in + fan_out)).
+    is
 
-    References:
-        [1] Understanding the difficulty of training deep feedforward neural
-            networks. International conference on artificial intelligence and
-            statistics.
-            (http://proceedings.mlr.press/v9/glorot10a.html)
-    """
+    .. math::
+
+        \sqrt{\\frac{2.0}{fan\_in + fan\_out}}
 
-    def __init__(self, uniform=True, fan_in=None, fan_out=None, seed=0):
-        """Constructor for XavierInitializer
 
     Args:
-            uniform: whether to use uniform or normal distribution
-            fan_in: fan_in for Xavier initialization. If None, it is
+        uniform (bool): whether to use uniform or normal distribution
+        fan_in (float): fan_in for Xavier initialization. If None, it is
                 inferred from the variable.
-            fan_out: fan_out for Xavier initialization. If None, it is
+        fan_out (float): fan_out for Xavier initialization. If None, it is
                  inferred from the variable.
-            seed: random seed
+        seed (int): random seed
+
+    Note:
+        It is recommended to set fan_in and fan_out to None for most cases.
+
+    Examples:
+        .. code-block:: python
+
+            fc = fluid.layers.fc(
+                input=queries, size=10,
+                param_attr=fluid.initializer.Xavier(uniform=False))
 
-        Note: It is recommended to set fan_in and fan_out to None for
-              most cases.
     """
+
+    def __init__(self, uniform=True, fan_in=None, fan_out=None, seed=0):
         assert uniform is not None
         assert seed is not None
         super(XavierInitializer, self).__init__()
@@ -342,30 +377,42 @@ class MSRAInitializer(Initializer):
     """Implements the MSRA initializer a.k.a. Kaiming Initializer
 
     This class implements the weight initialization from the paper
-    Delving Deep into Rectifiers: Surpassing Human-Level Performance on
-    ImageNet Classification[1] by Kaiming He, Xiangyu Zhang, Shaoqing Ren
-    and Jian Sun. This is a robust initialization method that particularly
-    considers the rectifier nonlinearities. In case of Uniform distribution,
-    the range is [-x, x], where x = sqrt(6 / fan_in). In case of Normal
-    distribution, the mean is 0 and the standard deviation
-    is sqrt(2/ fan_in).
-
-    References:
-        [1] Delving Deep into Rectifiers: Surpassing Human-Level Performance
-            on ImageNet Classification
-            (https://arxiv.org/abs/1502.01852)
-    """
+    `Delving Deep into Rectifiers: Surpassing Human-Level Performance on
+    ImageNet Classification <https://arxiv.org/abs/1502.01852>`_
+    by Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun. This is a
+    robust initialization method that particularly considers the rectifier
+    nonlinearities. In case of Uniform distribution, the range is [-x, x], where
 
-    def __init__(self, uniform=True, fan_in=None, seed=0):
-        """Constructor for MSRAInitializer
+    .. math::
+
+        x = \sqrt{\\frac{6.0}{fan\_in}}
+
+    In case of Normal distribution, the mean is 0 and the standard deviation
+    is
+
+    .. math::
+
+        \sqrt{\\frac{2.0}{fan\_in}}
 
     Args:
-            uniform: whether to use uniform or normal distribution
-            fan_in: fan_in for MSRAInitializer. If None, it is
+        uniform (bool): whether to use uniform or normal distribution
+        fan_in (float): fan_in for MSRAInitializer. If None, it is\
         inferred from the variable.
-            seed: random seed
+        seed (int): random seed
+
+    Note:
+        It is recommended to set fan_in to None for most cases.
+
+    Examples:
+        .. code-block:: python
+
+            fc = fluid.layers.fc(
+                input=queries, size=10,
+                param_attr=fluid.initializer.MSRA(uniform=False))
+    """
 
-        Note: It is recommended to set fan_in to None for most cases.
+    def __init__(self, uniform=True, fan_in=None, seed=0):
+        """Constructor for MSRAInitializer
         """
         assert uniform is not None
         assert seed is not None
@@ -425,34 +472,37 @@ class MSRAInitializer(Initializer):
 
 
 class BilinearInitializer(Initializer):
-    """Implements the bilinear initializer.
-
+    """
     This initializer can be used in transposed convolution operator to
     act as upsampling. Users can upsample a feature map with shape of
     (B, C, H, W) by any integer factor. The usage is:
 
-    >>>  factor = 2
-    >>>  w_attr = ParamAttr(learning_rate=0., regularizer=L2Decay(0.),
-    >>>                     initializer=Bilinear())
-    >>>  conv_up = fluid.layers.conv2d_transpose(
-    >>>      input,
-    >>>      num_filters=C,
-    >>>      output_size=None,
-    >>>      filter_size=2 * factor - factor % 2,
-    >>>      padding=ceil((factor - 1) / 2.),
-    >>>      stride=factor,
-    >>>      groups=C,
-    >>>      param_attr=w_attr,
-    >>>      bias_attr=False)
-
-
-    Where, `num_filters=C` and `groups=C` means this is channel-wise tranposed
+    Examples:
+
+        .. code-block:: python
+
+            factor = 2
+            w_attr = ParamAttr(learning_rate=0., regularizer=L2Decay(0.),
+                               initializer=Bilinear())
+            conv_up = fluid.layers.conv2d_transpose(
+                input,
+                num_filters=C,
+                output_size=None,
+                filter_size=2 * factor - factor % 2,
+                padding=ceil((factor - 1) / 2.),
+                stride=factor,
+                groups=C,
+                param_attr=w_attr,
+                bias_attr=False)
+
+    Where, `num_filters=C` and `groups=C` means this is channel-wise transposed
     convolution. The filter shape will be (C, 1, K, K) where K is `filer_size`,
     This initializer will set a (K, K) interpolation kernel for every channel
     of the filter identically. The resulting shape of the output feature map
     will be (B, C, factor * H, factor * W). Note that the learning rate and the
     weight decay are set to 0 in order to keep coefficient values of bilinear
     interpolation unchanged during training.
+
     """
 
     def __init__(self):
@@ -469,7 +519,7 @@ class BilinearInitializer(Initializer):
                            be added.
 
         Returns:
-            the initialization op
+            Operator: the initialization op
 
         Raises:
             ValueError: If type of `var` and `block` is not right.

python/paddle/fluid/layers/control_flow.py

@@ -185,7 +185,9 @@ def Print(input,
     Returns:
         Variable: Output tensor, same data with input tensor.
 
+
     Examples:
+
         .. code-block:: python
 
            value = some_layer(...)
@@ -1201,6 +1203,31 @@ class ConditionalBlockGuard(BlockGuard):
 
 
 class ConditionalBlock(object):
+    '''
+    **ConditionalBlock**
+
+    ConditionalBlock is an operator that bind a block to a specific condition,
+    if the condition matches, the corresponding block will be executed.
+
+    Args:
+        inputs (Variable): bool conditions.
+        is_scalar_condition (bool): whether the branch is controled by a scalar.
+        name(str): name of this ConditionalBlock.
+
+    Examples:
+        .. code-block:: python
+
+             cond = layers.less_than(x=label, y=limit)
+             true_image, false_image = layers.split_lod_tensor(
+                 input=image, mask=cond)
+             true_cond = layers.ConditionalBlock([true_image])
+
+             with true_cond.block():
+                 ...
+             with false_cond.block():
+                 ...
+    '''
+
     def __init__(self, inputs, is_scalar_condition=False, name=None):
         for each_input in inputs:
             if not isinstance(each_input, Variable):

python/paddle/fluid/layers/detection.py

@@ -16,7 +16,7 @@ All layers just related to the detection neural network.
 """
 
 from layer_function_generator import generate_layer_fn
-from layer_function_generator import autodoc
+from layer_function_generator import autodoc, templatedoc
 from ..layer_helper import LayerHelper
 import tensor
 import nn
@@ -155,7 +155,7 @@ def detection_output(loc,
     return nmsed_outs
 
 
-@autodoc()
+@templatedoc()
 def detection_map(detect_res,
                   label,
                   class_num,
@@ -166,6 +166,47 @@ def detection_map(detect_res,
                   input_states=None,
                   out_states=None,
                   ap_version='integral'):
+    """
+    ${comment}
+
+    Args:
+        detect_res: ${detect_res_comment}
+        label:  ${label_comment}
+        class_num: ${class_num_comment}
+        background_label: ${background_label_comment}
+        overlap_threshold: ${overlap_threshold_comment}
+        evaluate_difficult: ${evaluate_difficult_comment}
+        has_state: ${has_state_comment}
+        input_states: If not None, It contains 3 elements:
+            1. pos_count ${pos_count_comment}.
+            2. true_pos ${true_pos_comment}.
+            3. false_pos ${false_pos_comment}.
+        out_states: If not None, it contains 3 elements.
+            1. accum_pos_count ${accum_pos_count_comment}.
+            2. accum_true_pos ${accum_true_pos_comment}.
+            3. accum_false_pos ${accum_false_pos_comment}.
+        ap_version: ${ap_type_comment}
+
+    Returns:
+        ${map_comment}
+
+
+    Examples:
+          .. code-block:: python
+
+            detect_res = fluid.layers.data(
+                name='detect_res',
+                shape=[10, 6],
+                append_batch_size=False,
+                dtype='float32')
+            label = fluid.layers.data(
+                name='label',
+                shape=[10, 6],
+                append_batch_size=False,
+                dtype='float32')
+
+            map_out = fluid.layers.detection_map(detect_res, label, 21)
+    """
     helper = LayerHelper("detection_map", **locals())
 
     def __create_var(type):

python/paddle/fluid/layers/nn.py

@@ -93,6 +93,7 @@ __all__ = [
     'mean_iou',
     'relu',
     'log',
+    'crop',
 ]
 
 
@@ -197,7 +198,10 @@ def fc(input,
     else:
         pre_bias = helper.create_tmp_variable(dtype)
         helper.append_op(
-            type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias})
+            type="sum",
+            inputs={"X": mul_results},
+            outputs={"Out": pre_bias},
+            attrs={"use_mkldnn": use_mkldnn})
     # add bias
     pre_activation = helper.append_bias_op(pre_bias, dim_start=num_flatten_dims)
     # add activation
@@ -2674,18 +2678,35 @@ def sequence_expand(x, y, ref_level=-1, name=None):
 
 def beam_search(pre_ids, ids, scores, beam_size, end_id, level=0):
     '''
+    **beam search**
+
     This function implements the beam search algorithm.
 
+    Beam search is a classical algorithm for selecting candidate words
+    in a machine translation task.
+
+    Refer to `Beam search <https://en.wikipedia.org/wiki/Beam_search>`_
+    for more details.
+
     Args:
-        pre_ids (Variable): ${pre_ids_comment}
-        ids (Variable): ${ids_comment}
-        scores (Variable): ${scores_comment}
-        beam_size (int): ${beam_size_comment}
-        end_id (int): ${end_id_comment}
-        level (int): ${level_comment}
+        pre_ids (Variable): ids in previous step.
+        ids (Variable): a LoDTensor of shape of [None,k]
+        scores (Variable): a LoDTensor that has the same shape and LoD with `ids`
+        beam_size (int): beam size for beam search
+        end_id (int): the token id which indicates the end of a sequence
+        level (int): the level of LoDTensor
 
     Returns:
-        tuple: a tuple of beam_search output variables: selected_ids, selected_scores
+        tuple: a tuple of beam_search output variables: `selected_ids`, `selected_scores`
+
+    Examples:
+        .. code-block:: python
+
+             # current_score is a Tensor of shape (num_batch_size, embed_size), which
+             # consists score of each candidate word.
+             topk_scores, topk_indices = pd.topk(current_score, k=50)
+             selected_ids, selected_scores = pd.beam_search(
+                 pre_ids, topk_indices, topk_scores, beam_size, end_id=10, level=0)
     '''
     helper = LayerHelper('beam_search', **locals())
     score_type = scores.dtype
@@ -5003,3 +5024,101 @@ def mean_iou(input, label, num_classes):
         },
         attrs={"num_classes": num_classes})
     return out_mean_iou, out_wrong, out_correct
+
+
+def crop(x, shape=None, offsets=None, name=None):
+    """
+    Crop input into output, as specified by offsets and shape.
+
+    .. code-block:: text
+
+        * Case 1:
+            Given
+                X = [[0, 1, 2, 0, 0]
+                     [0, 3, 4, 0, 0]
+                     [0, 0, 0, 0, 0]],
+            and
+                shape = [2, 2],
+                offsets = [0, 1],
+            output is:
+                Out = [[1, 2],
+                       [3, 4]].
+        * Case 2:
+            Given
+                X = [[0, 1, 2, 5, 0]
+                     [0, 3, 4, 6, 0]
+                     [0, 0, 0, 0, 0]],
+            and shape is tensor
+                shape = [[0, 0, 0]
+                         [0, 0, 0]]
+            and
+                offsets = [0, 1],
+
+            output is:
+                Out = [[1, 2, 5],
+                       [3, 4, 6]].
+
+    Args:
+        x (Variable): The input tensor variable.
+        shape (Variable|list/tuple of integer): The output shape is specified
+            by `shape`, which can a Variable or a list/tupe of integer.
+            If a tensor Variable, it's rank must be the same as `x`. This way
+            is suitable for the case that the output shape may be changed each
+            iteration. If a list/tupe of integer, it's length must be the same
+            as the rank of `x`
+        offsets (Variable|list/tuple of integer|None): Specifies the copping
+            offsets at each dimension. It can be a Variable or or a list/tupe
+            of integer. If a tensor Variable, it's rank must be the same as `x`.
+            This way is suitable for the case that the offsets may be changed
+            each iteration. If a list/tupe of integer, it's length must be the
+            same as the rank of `x`. If None, the offsets are 0 at each
+            dimension.
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
+
+    Returns:
+        Variable: The cropped tensor variable.
+
+    Raises:
+        ValueError: If shape is not a list, tuple or Variable.
+
+    Examples:
+
+        .. code-block:: python
+
+            x = fluid.layers.data(name="x", shape=[3, 5], dtype="float32")
+            y = fluid.layers.data(name="y", shape=[2, 3], dtype="float32")
+            crop = fluid.layers.crop(x, shape=y)
+
+            # or
+            z = fluid.layers.data(name="z", shape=[3, 5], dtype="float32")
+            crop = fluid.layers.crop(z, shape=[2, 3])
+
+    """
+    helper = LayerHelper('crop', **locals())
+
+    if not (isinstance(shape, list) or isinstance(shape, tuple) or \
+        isinstance(shape, Variable)):
+        raise ValueError("The shape should be a list, tuple or Variable.")
+
+    if offsets is None:
+        offsets = [0] * len(x.shape)
+
+    out = helper.create_tmp_variable(x.dtype)
+    ipts = {'X': x}
+    attrs = {}
+    if isinstance(shape, Variable):
+        ipts['Y'] = shape
+    else:
+        attrs['shape'] = shape
+    if isinstance(offsets, Variable):
+        ipts['Offsets'] = offsets
+    else:
+        attrs['offsets'] = offsets
+
+    helper.append_op(
+        type='crop',
+        inputs=ipts,
+        outputs={'Out': out},
+        attrs=None if len(attrs) == 0 else attrs)
+    return out

python/paddle/fluid/layers/tensor.py

@@ -230,7 +230,11 @@ def sums(input, out=None):
     helper = LayerHelper('sum', **locals())
     if out is None:
         out = helper.create_tmp_variable(dtype=helper.input_dtype())
-    helper.append_op(type='sum', inputs={'X': input}, outputs={'Out': out})
+        helper.append_op(
+            type='sum',
+            inputs={'X': input},
+            outputs={'Out': out},
+            attrs={'use_mkldnn': False})
     return out
 
 

python/paddle/fluid/optimizer.py

@@ -26,10 +26,10 @@ from clip import append_gradient_clip_ops, error_clip_callback
 from contextlib import contextmanager
 
 __all__ = [
-    'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad',
+    'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'Ftrl',
     'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer',
     'AdamaxOptimizer', 'DecayedAdagradOptimizer', 'RMSPropOptimizer',
-    'Adadelta', 'ModelAverage', 'Optimizer'
+    'FtrlOptimizer', 'Adadelta', 'ModelAverage', 'Optimizer', 'RMSPropOptimizer'
 ]
 
 
@@ -192,15 +192,15 @@ class Optimizer(object):
         """Add optimization operators to update gradients to variables.
 
         Args:
-          loss: the target that this optimization is for.
-          parameters_and_grads: a list of (variable, gradient) pair to update.
+          loss(Variable): the target that this optimization is for.
+          parameters_and_grads(list(tuple(Variable, Variable))):
+          a list of (variable, gradient) pair to update.
 
         Returns:
           return_op_list: a list of operators that will complete one step of
           optimization. This will include parameter update ops, global step
           update ops and any other custom ops required by subclasses to manage
           their internal state.
-          :param startup_program:
         """
         # This is a default implementation of create_optimization_pass that
         # can be shared by most optimizers. This implementation assumes that
@@ -268,7 +268,22 @@ class Optimizer(object):
 
 
 class SGDOptimizer(Optimizer):
-    """ Simple SGD optimizer without any state.
+    """
+    Optimizer of the stochastic gradient descent algorithm.
+
+    .. math::
+
+        param\_out = param - learning\_rate * grad
+
+    Args:
+        learning_rate (float|Variable): the learning rate used to update parameters. \
+        Can be a float value or a Variable with one float value as data element.
+
+    Examples:
+        .. code-block:: python
+
+            sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.2)
+            sgd_optimizer.minimize(cost)
     """
 
     def __init__(self, learning_rate, **kwargs):
@@ -294,7 +309,37 @@ class SGDOptimizer(Optimizer):
 
 
 class MomentumOptimizer(Optimizer):
-    """Simple Momentum optimizer with velocity state
+    """
+
+    Simple Momentum optimizer with velocity state
+
+    This optimizer has a flag for Nestrov Momentum.
+
+    The update equations are as follows:
+
+    .. math::
+
+        & velocity = mu * velocity + gradient
+
+        & if (use\_nesterov):
+
+        &\quad   param = param - gradient * learning\_rate + mu * velocity * learning\_rate
+
+        & else:
+
+        &\quad   param = param - learning\_rate * velocity
+
+    Args:
+        learning_rate (float|Variable): the learning rate used to update parameters. \
+        Can be a float value or a Variable with one float value as data element.
+        momentum (float): momentum factor
+        use_nesterov (bool): enables Nesterov momentum
+
+    Examples:
+        .. code-block:: python
+
+            optimizer = fluid.optimizer.Momentum(learning_rate=0.2, momentum=0.1)
+            optimizer.minimize(cost)
     """
     _velocity_acc_str = "velocity"
 
@@ -338,7 +383,32 @@ class MomentumOptimizer(Optimizer):
 
 
 class AdagradOptimizer(Optimizer):
-    """Simple Adagrad optimizer with moment state
+    """
+    **Adaptive Gradient Algorithm (Adagrad)**
+
+    The update is done as follows:
+
+    .. math::
+
+        moment\_out &= moment + grad * grad
+
+        param\_out &= param - \\frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon}
+
+    The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
+    does not have the epsilon attribute. It is added here in our implementation
+    as also proposed here: http://cs231n.github.io/neural-networks-3/#ada
+    for numerical stability to avoid the division by zero error.
+
+    Args:
+        learning_rate (float|Variable): the learning rate used to update parameters. \
+        Can be a float value or a Variable with one float value as data element.
+        epsilon (float): a small float value for numerical stability.
+
+    Examples:
+        .. code-block:: python
+
+            optimizer = fluid.optimizer.Adagrad(learning_rate=0.2)
+            optimizer.minimize(cost)
     """
     _moment_acc_str = "moment"
 
@@ -379,7 +449,40 @@ class AdagradOptimizer(Optimizer):
 
 
 class AdamOptimizer(Optimizer):
-    """Implements the Adam Optimizer
+    """
+    This implements the Adam optimizer from Section 2 of the Adam
+    paper : https://arxiv.org/abs/1412.6980.
+    Adam is a first-order gradient-based optimization method based on
+    adaptive estimates of lower-order moments.
+
+    Adam updates:
+
+    .. math::
+
+        t & = t + 1
+
+        moment\_1\_out & = {\\beta}_1 * moment\_1 + (1 - {\\beta}_1) * grad
+
+        moment\_2\_out & = {\\beta}_2 * moment\_2 + (1 - {\\beta}_2) * grad * grad
+
+        learning\_rate & = learning\_rate * \\
+                          \\frac{\sqrt{1 - {\\beta}_2^t}}{1 - {\\beta}_1^t}
+
+        param\_out & = param - learning\_rate * \\frac{moment\_1}{\sqrt{moment\_2} + \epsilon}
+
+    Args:
+        learning_rate (float|Variable): the learning rate used to update parameters. \
+        Can be a float value or a Variable with one float value as data element.
+        beta1 (float): The exponential decay rate for the 1st moment estimates.
+        beta2 (float): The exponential decay rate for the 2nd moment estimates.
+        epsilon (float): a small float value for numerical stability.
+
+    Examples:
+        .. code-block:: python
+
+            optimizer = fluid.optimizer.Adam(learning_rate=0.2)
+            optimizer.minimize(cost)
+
     """
     _moment1_acc_str = "moment1"
     _moment2_acc_str = "moment2"
@@ -484,7 +587,42 @@ class AdamOptimizer(Optimizer):
 
 
 class AdamaxOptimizer(Optimizer):
-    """Implements the Adamax Optimizer
+    """
+    We implement the Adamax optimizer from Section 7 of the Adam
+    paper: https://arxiv.org/abs/1412.6980. Adamax is a variant of the
+    Adam algorithm based on the infinity norm.
+
+    Adamax updates:
+
+    .. math::
+
+        t & = t + 1
+
+        moment\_out & = {\\beta}_1 * moment + (1 - {\\beta}_1) * grad
+
+        inf\_norm\_out & = max({\\beta}_2 * inf\_norm + \epsilon, |grad|)
+
+        learning\_rate & = \\frac{learning\_rate}{1 - {\\beta}_1^t}
+
+        param\_out & = param - learning\_rate * \\frac{moment\_out}{inf\_norm\_out}
+
+
+    The original paper does not have an epsilon attribute.
+    However, it is added here for numerical stability to prevent the
+    division by 0 error.
+
+    Args:
+        learning_rate (float|Variable): the learning rate used to update parameters. \
+        Can be a float value or a Variable with one float value as data element.
+        beta1 (float): The exponential decay rate for the 1st moment estimates.
+        beta2 (float): The exponential decay rate for the 2nd moment estimates.
+        epsilon (float): a small float value for numerical stability.
+
+    Examples:
+        .. code-block:: python
+
+            optimizer = fluid.optimizer.Adamax(learning_rate=0.2)
+            optimizer.minimize(cost)
     """
     _moment_acc_str = "moment"
     _inf_norm_acc_str = "inf_norm"
@@ -568,7 +706,34 @@ class AdamaxOptimizer(Optimizer):
 
 
 class DecayedAdagradOptimizer(Optimizer):
-    """Simple Decayed Adagrad optimizer with moment state
+    """
+    **Decayed Adagrad Optimizer**
+
+    The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
+
+    The update is done as follows:
+
+    .. math::
+
+        moment\_out & = decay * moment + (1 - decay) * grad * grad
+
+        param\_out & = param - \\frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon}
+
+    The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
+    does not have an epsilon attribute. It is added here for numerical
+    stability to avoid the division by zero error.
+
+    Args:
+        learning_rate (float|Variable): the learning rate used to update parameters. \
+        Can be a float value or a Variable with one float value as data element.
+        decay (float): decay rate.
+        epsilon (float): a small float value for numerical stability.
+
+    Examples:
+        .. code-block:: python
+
+            optimizer = fluid.optimizer.DecayedAdagrad(learning_rate=0.2)
+            optimizer.minimize(cost)
     """
     _moment_acc_str = "moment"
 
@@ -614,6 +779,7 @@ class DecayedAdagradOptimizer(Optimizer):
 class AdadeltaOptimizer(Optimizer):
     """
     **Adadelta Optimizer**
+
     Simple Adadelta optimizer with average squared grad state and
     average squared update state.
     The details of adadelta please refer to this
@@ -628,7 +794,7 @@ class AdadeltaOptimizer(Optimizer):
         E(dx_t^2) &= \\rho * E(dx_{t-1}^2) + (1-\\rho) * (-g*learning\\_rate)^2
 
     Args:
-        learning_rate(float): global leraning rate
+        learning_rate(float): global learning rate
         rho(float): rho in equation
         epsilon(float): epsilon in equation
 
@@ -703,37 +869,37 @@ class RMSPropOptimizer(Optimizer):
 
     ..  math::
 
-        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 \\\\
+        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
 
         w & = w - \\frac{\\eta} {\\sqrt{r(w,t) + \\epsilon}} \\nabla Q_{i}(w)
 
     The first equation calculates moving average of the squared gradient for
-    each weight. Then dividing the gradient by :math: `sqrt{v(w,t)}`.
+    each weight. Then dividing the gradient by :math:`sqrt{v(w,t)}`.
 
     In some cases, adding a momentum term :math: `\\beta` is beneficial.
     In our implementation, Nesterov momentum is used:
 
     ..  math::
 
-        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 \\\\
+        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
 
         v(w, t) & = \\beta v(w, t-1) + \\frac{\\eta} {\\sqrt{v(w,t) +
             \\epsilon}} \\nabla Q_{i}(w)
 
         w & = w - v(w, t)
 
-    where, :math: `\\rho` is a hyperparameter and typical values are 0.9, 0.95
+    where, :math:`\\rho` is a hyperparameter and typical values are 0.9, 0.95
     and so on. :math: `beta` is the momentum term. :math: `\\epsilon` is a
     smoothing term to avoid division by zero, usually set somewhere in range
     from 1e-4 to 1e-8.
 
 
     Args:
-        learning_rate(float): global leraning rate.
+        learning_rate(float): global learning rate.
         rho(float): rho is :math: `\\rho` in equation, set 0.95 by default.
         epsilon(float): :math: `\\epsilon` in equation is smoothing term to
             avoid division by zero, set 1e-6 by default.
-        momentum(float): :math: `\\beta` in equation is the momentum term,
+        momentum(float): :math:`\\beta` in equation is the momentum term,
             set 0.0 by default.
 
     Raises:
@@ -810,6 +976,113 @@ class RMSPropOptimizer(Optimizer):
         return rmsprop_op
 
 
+class FtrlOptimizer(Optimizer):
+    """
+    FTRL (Follow The Regularized Leader) Optimizer.
+
+    The paper that proposed Follow The Regularized Leader (FTRL):
+    (https://www.eecs.tufts.edu/~dsculley/papers/ad-click-prediction.pdf)
+
+    ..  math::
+
+        &new\_accum = squared\_accum + grad^2
+
+        &if (lr\_power == -0.5):
+
+        &\quad  linear\_accum += grad - \\frac{\\sqrt{new\_accum} - \\sqrt{squared\_accum}}{learning\_rate * param}
+
+        &else:
+
+        &\quad   linear\_accum += grad - \\frac{new\_accum^{-lr\_power} - accum^{-lr\_power}}{learning\_rate * param}
+
+
+        &x = l1 * sign(linear\_accum) - linear\_accum
+
+        &if (lr\_power == -0.5):
+
+        &\quad   y = \\frac{\\sqrt{new\_accum}}{learning\_rate} + (2 * l2)
+
+        &\quad   pre\_shrink = \\frac{x}{y}
+
+        &\quad   param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0)
+
+        &else:
+
+        &\quad   y = \\frac{new\_accum^{-lr\_power}}{learning\_rate} + (2 * l2)
+
+        &\quad   pre\_shrink = \\frac{x}{y}
+
+        &\quad   param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0)
+
+        &squared\_accum += grad^2
+
+    Args:
+        learning_rate (float|Variable): global learning rate.
+        l1 (float):
+        l2 (float):
+        lr_power (float):
+
+    Raises:
+        ValueError: If learning_rate, rho, epsilon, momentum are None.
+
+    Examples:
+          .. code-block:: python
+
+              optimizer = fluid.optimizer.Ftrl(0.0001)
+              _, params_grads = optimizer.minimize(cost)
+    """
+
+    _squared_acc_str = "squared"
+    _linear_acc_str = "linear"
+
+    def __init__(self, learning_rate, l1=0.0, l2=0.0, lr_power=-0.5, **kwargs):
+        super(FtrlOptimizer, self).__init__(
+            learning_rate=learning_rate, **kwargs)
+        if learning_rate is None:
+            raise ValueError("learning_rate is not set.")
+
+        self.type = "ftrl"
+        self._l1 = l1
+        self._l2 = l2
+        self._lr_power = lr_power
+
+    def _create_accumulators(self, block, parameters):
+        if not isinstance(block, framework.Block):
+            raise TypeError("block is not instance of framework.Block.")
+
+        for p in parameters:
+            self._add_accumulator(self._squared_acc_str, p)
+            self._add_accumulator(self._linear_acc_str, p)
+
+    def _append_optimize_op(self, block, param_and_grad):
+        if not isinstance(block, framework.Block):
+            raise TypeError("block is not instance of framework.Block.")
+
+        squared_acc = self._get_accumulator(self._squared_acc_str,
+                                            param_and_grad[0])
+        linear_acc = self._get_accumulator(self._linear_acc_str,
+                                           param_and_grad[0])
+        ftrl_op = block.append_op(
+            type=self.type,
+            inputs={
+                "Param": param_and_grad[0],
+                "Grad": param_and_grad[1],
+                "SquaredAccumulator": squared_acc,
+                "LinearAccumulator": linear_acc,
+                "LearningRate": self._create_param_lr(param_and_grad),
+            },
+            outputs={
+                "ParamOut": param_and_grad[0],
+                "SquaredAccumOut": squared_acc,
+                "LinearAccumOut": linear_acc
+            },
+            attrs={"l1": self._l1,
+                   "l2": self._l1,
+                   "lr_power": self._lr_power})
+
+        return ftrl_op
+
+
 # We short the class name, since users will use the optimizer with the package
 # name. The sample code:
 #
@@ -826,6 +1099,7 @@ Adamax = AdamaxOptimizer
 DecayedAdagrad = DecayedAdagradOptimizer
 Adadelta = AdadeltaOptimizer
 RMSProp = RMSPropOptimizer
+Ftrl = FtrlOptimizer
 
 
 class ModelAverage(Optimizer):
@@ -844,7 +1118,9 @@ class ModelAverage(Optimizer):
         max_average_window: The maximum size of average window.
 
     Examples:
-        ...
+
+      .. code-block:: python
+
         optimizer = fluid.optimizer.Momentum()
         _, params_grads = optimizer.minimize(cost)
         model_average = fluid.optimizer.ModelAverage(params_grads, 0.15,

python/paddle/fluid/profiler.py

@@ -42,6 +42,9 @@ def cuda_profiler(output_file, output_mode=None, config=None):
     counters/options for profiling by `config` argument. The default config
     is ['gpustarttimestamp', 'gpustarttimestamp', 'gridsize3d',
     'threadblocksize', 'streamid', 'enableonstart 0', 'conckerneltrace'].
+    Then users can use NVIDIA Visual Profiler
+    (https://developer.nvidia.com/nvidia-visual-profiler) tools to load this
+    this output file to visualize results.
 
     Args:
         output_file (string) : The output file name, the result will be
@@ -50,6 +53,33 @@ def cuda_profiler(output_file, output_mode=None, config=None):
             Comma separated values format. It should be 'kvp' or 'csv'.
         config (list of string) : The profiler options and counters can refer
             to "Compute Command Line Profiler User Guide".
+
+    Raises:
+        ValueError: If `output_mode` is not in ['kvp', 'csv'].
+
+    Examples:
+
+        .. code-block:: python
+
+            import paddle.fluid as fluid
+            import paddle.fluid.profiler as profiler
+
+            epoc = 8
+            dshape = [4, 3, 28, 28]
+            data = fluid.layers.data(name='data', shape=[3, 28, 28], dtype='float32')
+            conv = fluid.layers.conv2d(data, 20, 3, stride=[1, 1], padding=[1, 1])
+
+            place = fluid.CUDAPlace(0)
+            exe = fluid.Executor(place)
+            exe.run(fluid.default_startup_program())
+
+            output_file = 'cuda_profiler.txt'
+            with profiler.cuda_profiler(output_file, 'csv') as nvprof:
+                for i in range(epoc):
+                    input = np.random.random(dshape).astype('float32')
+                    exe.run(fluid.default_main_program(), feed={'data': input})
+            # then use  NVIDIA Visual Profiler (nvvp) to load this output file
+            # to visualize results.
     """
     if output_mode is None:
         output_mode = 'csv'
@@ -69,19 +99,52 @@ def cuda_profiler(output_file, output_mode=None, config=None):
 
 
 def reset_profiler():
-    """The profiler clear interface.
-    reset_profiler will clear the previous time record.
+    """
+    Clear the previous time record. This interface does not work for
+    `fluid.profiler.cuda_profiler`, it only works for
+    `fluid.profiler.start_profiler`, `fluid.profiler.stop_profiler`,
+    and `fluid.profiler.profiler`.
+
+    Examples:
+
+        .. code-block:: python
+
+            import paddle.fluid.profiler as profiler
+            with profiler.profiler(state, 'total', '/tmp/profile'):
+                for iter in range(10):
+                    if iter == 2:
+                        profiler.reset_profiler()
+                    # ...
     """
     core.reset_profiler()
 
 
 def start_profiler(state):
-    """Enable the profiler.
+    """
+    Enable the profiler. Uers can use `fluid.profiler.start_profiler` and
+    `fluid.profiler.stop_profiler` to insert the code, except the usage of
+    `fluid.profiler.profiler` interface.
 
     Args:
         state (string) : The profiling state, which should be 'CPU', 'GPU'
             or 'All'. 'CPU' means only profile CPU. 'GPU' means profiling
             GPU as well. 'All' also generates timeline.
+
+    Raises:
+        ValueError: If `state` is not in ['CPU', 'GPU', 'All'].
+
+    Examples:
+
+        .. code-block:: python
+
+            import paddle.fluid.profiler as profiler
+
+            profiler.start_profiler('GPU')
+            for iter in range(10):
+                if iter == 2:
+                    profiler.reset_profiler()
+                # except each iteration
+            profiler.stop_profiler('total', '/tmp/profile')
     """
     if core.is_profiler_enabled():
         return
@@ -97,7 +160,10 @@ def start_profiler(state):
 
 
 def stop_profiler(sorted_key=None, profile_path='/tmp/profile'):
-    """Stop the profiler.
+    """
+    Stop the profiler. Uers can use `fluid.profiler.start_profiler` and
+    `fluid.profiler.stop_profiler` to insert the code, except the usage of
+    `fluid.profiler.profiler` interface.
 
     Args:
         sorted_key (string) : If None, the profiling results will be printed
@@ -111,6 +177,23 @@ def stop_profiler(sorted_key=None, profile_path='/tmp/profile'):
             The `ave` means sorting by the average execution time.
         profile_path (string) : If state == 'All', it will write a profile
             proto output file.
+
+    Raises:
+        ValueError: If `sorted_key` is not in
+            ['calls', 'total', 'max', 'min', 'ave'].
+
+    Examples:
+
+        .. code-block:: python
+
+            import paddle.fluid.profiler as profiler
+
+            profiler.start_profiler('GPU')
+            for iter in range(10):
+                if iter == 2:
+                    profiler.reset_profiler()
+                # except each iteration
+            profiler.stop_profiler('total', '/tmp/profile')
     """
     if not core.is_profiler_enabled():
         return
@@ -137,7 +220,12 @@ def profiler(state, sorted_key=None, profile_path='/tmp/profile'):
     Different from cuda_profiler, this profiler can be used to profile both CPU
     and GPU program. By defalut, it records the CPU and GPU operator kernels,
     if you want to profile other program, you can refer the profiling tutorial
-    to add more records.
+    to add more records in C++ code.
+
+    If the state == 'All', a profile proto file will be written to
+    `profile_path`. This file records timeline information during the execution.
+    Then users can visualize this file to see the timeline, please refer 
+    https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/howto/optimization/timeline.md
 
     Args:
         state (string) : The profiling state, which should be 'CPU' or 'GPU',
@@ -156,6 +244,25 @@ def profiler(state, sorted_key=None, profile_path='/tmp/profile'):
             The `ave` means sorting by the average execution time.
         profile_path (string) : If state == 'All', it will write a profile
             proto output file.
+
+    Raises:
+        ValueError: If `state` is not in ['CPU', 'GPU', 'All']. If `sorted_key` is
+            not in ['calls', 'total', 'max', 'min', 'ave'].
+
+    Examples:
+
+        .. code-block:: python
+
+            import paddle.fluid.profiler as profiler
+
+            with profiler.profiler('All', 'total', '/tmp/profile') as prof:
+                for pass_id in range(pass_num):
+                    for batch_id, data in enumerate(train_reader()):
+                        exe.run(fluid.default_main_program(),
+                                feed=feeder.feed(data),
+                                fetch_list=[],
+                                use_program_cache=True)
+                        # ...
     """
     start_profiler(state)
     yield

python/paddle/fluid/regularizer.py

@@ -16,8 +16,8 @@ import framework
 from . import core
 
 __all__ = [
-    'append_regularization_ops', 'WeightDecayRegularizer', 'L1Decay', 'L2Decay',
-    'L1DecayRegularizer', 'L2DecayRegularizer'
+    'append_regularization_ops', 'L1Decay', 'L2Decay', 'L1DecayRegularizer',
+    'L2DecayRegularizer'
 ]
 
 
@@ -36,7 +36,8 @@ def append_regularization_ops(parameters_and_grads, regularization=None):
                         set. It will be applied with regularizer.
 
     Returns:
-        list of (parameters, gradients) pair with the regularized gradient
+        list[(Variable, Variable)]: list of (parameters, gradients) \
+        pair with the regularized gradient
 
     Raises:
         Exception: Unknown regularization type
@@ -100,6 +101,24 @@ class WeightDecayRegularizer(object):
 
 class L2DecayRegularizer(WeightDecayRegularizer):
     """Implements the L2 Weight Decay Regularization
+
+    Small values of L2 can help prevent over fitting the training data.
+
+    .. math::
+
+        L2WeightDecay = reg\_coeff * parameter
+
+    Args:
+        regularization_coeff(float): regularization coeff
+
+    Examples:
+        .. code-block:: python
+
+            optimizer = fluid.optimizer.Adagrad(
+                learning_rate=1e-4,
+                regularization=fluid.regularizer.L2DecayRegularizer(
+                    regularization_coeff=0.1))
+            optimizer.minimize(avg_cost)
     """
 
     def __init__(self, regularization_coeff=0.0):
@@ -154,6 +173,27 @@ class L2DecayRegularizer(WeightDecayRegularizer):
 
 class L1DecayRegularizer(WeightDecayRegularizer):
     """Implements the L1 Weight Decay Regularization
+
+    L1 regularization encourages sparsity.
+
+    .. math::
+
+        L1WeightDecay = reg\_coeff * sign(parameter)
+
+    Args:
+        regularization_coeff(float): regularization coeff
+
+    Examples:
+        .. code-block:: python
+
+            program = fluid.framework.Program()
+            block = program.global_block()
+            mul_x = block.create_parameter(
+                dtype="float32",
+                shape=[5, 10],
+                lod_level=0,
+                name="mul.x",
+                regularizer=fluid.regularizer.L1DecayRegularizer(0.5))
     """
 
     def __init__(self, regularization_coeff=0.0):

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

@@ -194,16 +194,16 @@ def train(word_dict,
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

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

@@ -69,16 +69,16 @@ def train(use_cuda, save_dirname, is_local):
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

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

@@ -178,16 +178,16 @@ def train(net_type, use_cuda, save_dirname, is_local):
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

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

@@ -209,16 +209,16 @@ def train(use_cuda, save_dirname=None, is_local=True):
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

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

@@ -200,16 +200,16 @@ def train_main(use_cuda, is_sparse, is_local=True):
     if is_local:
         train_loop(framework.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

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

@@ -151,16 +151,16 @@ def train(nn_type,
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

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

@@ -220,16 +220,16 @@ def train(use_cuda, save_dirname, is_local=True):
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

python/paddle/fluid/tests/book/test_word2vec.py

@@ -125,16 +125,16 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True):
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

python/paddle/fluid/tests/unittests/test_concat_op.py

@@ -43,7 +43,7 @@ class TestConcatOp(OpTest):
         self.axis = 1
 
 
-class TestConcatOp2(OpTest):
+class TestConcatOp2(TestConcatOp):
     def init_test_data(self):
         self.x0 = np.random.random((2, 3, 4, 5)).astype('float32')
         self.x1 = np.random.random((2, 3, 4, 5)).astype('float32')
@@ -51,5 +51,16 @@ class TestConcatOp2(OpTest):
         self.axis = 1
 
 
+class TestConcatOp3(TestConcatOp):
+    def init_test_data(self):
+        self.x0 = np.random.random((1, 256, 170, 256)).astype('float32')
+        self.x1 = np.random.random((1, 128, 170, 256)).astype('float32')
+        self.x2 = np.random.random((1, 128, 170, 256)).astype('float32')
+        self.axis = 1
+
+    def test_check_grad(self):
+        pass
+
+
 if __name__ == '__main__':
     unittest.main()

paddle/contrib/tape/CMakeLists.txt → python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py

-# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+# Copyright (c) 2018 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.
@@ -11,15 +11,16 @@
 # 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.
-#
 
-if(APPLE)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move")
-endif(APPLE)
+import unittest
+
+from test_gaussian_random_op import TestGaussianRandomOp
+
+
+class TestMKLDNN(TestGaussianRandomOp):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
 
-cc_library(tape_variable SRCS variable.cc DEPS ${FLUID_CORE_MODULES} device_context framework_proto proto_desc operator)
-cc_library(tape SRCS tape.cc DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB} tape_variable)
 
-cc_test(test_tape
-        SRCS test_tape.cc
-        DEPS tape tape_variable)
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_gaussian_random_op.py

@@ -25,7 +25,15 @@ class TestGaussianRandomOp(unittest.TestCase):
     def setUp(self):
         self.op_type = "gaussian_random"
         self.inputs = {}
-        self.attrs = {"shape": [1000, 784], "mean": .0, "std": 1., "seed": 10}
+        self.use_mkldnn = False
+        self.init_kernel_type()
+        self.attrs = {
+            "shape": [1000, 784],
+            "mean": .0,
+            "std": 1.,
+            "seed": 10,
+            "use_mkldnn": self.use_mkldnn
+        }
 
         self.outputs = ["Out"]
 
@@ -58,6 +66,9 @@ class TestGaussianRandomOp(unittest.TestCase):
         self.assertAlmostEqual(numpy.mean(tensor), .0, delta=0.1)
         self.assertAlmostEqual(numpy.std(tensor), 1., delta=0.1)
 
+    def init_kernel_type(self):
+        pass
+
 
 if __name__ == "__main__":
     unittest.main()

python/paddle/fluid/tests/unittests/test_layers.py

@@ -401,6 +401,15 @@ class TestBook(unittest.TestCase):
             self.assertIsNotNone(output)
         print(str(program))
 
+    def test_maxout(self):
+        program = Program()
+        with program_guard(program):
+            x = layers.data(name='x', shape=[3, 5], dtype="float32")
+            y = layers.data(name='y', shape=[2, 3], dtype="float32")
+            output = layers.crop(x, shape=y)
+            self.assertIsNotNone(output)
+        print(str(program))
+
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py

+# Copyright (c) 2018 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.
+
+import unittest
+
+from test_sum_op import TestSumOp
+
+
+class TestMKLDNN(TestSumOp):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+if __name__ == '__main__':
+    unittest.main()