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

doc/design/block.md

+# Design Doc: Block and Scope
+
+## The Representation of Computation
+
+Both deep learning systems and programming languages help users describe computation procedures.  These systems use various representations of computation:
+
+- Caffe, Torch, and Paddle: sequences of layers.
+- TensorFlow, Caffe2, Mxnet: graphs of operators.
+- PaddlePaddle: nested blocks, like C++ and Java programs.
+
+## Block in Programming Languages and Deep Learning
+
+In programming languages, a block is a pair of curly braces that includes local variables definitions and a sequence of instructions, or operators.
+
+Blocks work with control flow structures like `if`, `else`, and `for`, which have equivalents in deep learning:
+
+| programming languages | PaddlePaddle          |
+|-----------------------|-----------------------|
+| for, while loop       | RNN, WhileOp          |
+| if, if-else, switch   | IfElseOp, SwitchOp    |
+| sequential execution  | a sequence of layers  |
+
+A key difference is that a C++ program describes a one pass computation, whereas a deep learning program describes both the forward and backward passes.
+
+## Stack Frames and the Scope Hierarchy
+
+The existence of the backward makes the execution of a block of traditional programs and PaddlePaddle different to each other:
+
+| programming languages | PaddlePaddle                  |
+|-----------------------|-------------------------------|
+| stack                 | scope hierarchy               |
+| stack frame           | scope                         |
+| push at entering block| push at entering block        |
+| pop at leaving block  | destroy at minibatch completes|
+
+1. In traditional programs:
+
+   - When the execution enters the left curly brace of a block, the runtime pushes a frame into the stack, where it realizes local variables.
+   - After the execution leaves the right curly brace, the runtime pops the frame.
+   - The maximum number of frames in the stack is the maximum depth of nested blocks.
+
+1. In PaddlePaddle
+
+   - When the execution enters a block, PaddlePaddle adds a new scope, where it realizes variables.
+   - PaddlePaddle doesn't pop a scope after the execution of the block because variables therein are to be used by the backward pass.  So it has a stack forest known as a *scope hierarchy*.
+   - The height of the highest tree is the maximum depth of nested blocks.
+   - After the process of a minibatch, PaddlePaddle destroys the scope hierarchy.
+
+## Use Blocks in C++ and PaddlePaddle Programs
+
+Let us consolidate the discussion by presenting some examples.
+
+### Blocks with `if-else` and `IfElseOp`
+
+The following C++ programs shows how blocks are used with the `if-else` structure:
+
+```c++
+int x = 10;
+int y = 20;
+int out;
+bool cond = false;
+if (cond) {
+  int z = x + y;
+  out = softmax(z);
+} else {
+  int z = fc(x);
+  out = z;
+}
+```
+
+An equivalent PaddlePaddle program from the design doc of the [IfElseOp operator](./if_else_op.md) is as follows:
+
+```python
+import paddle as pd
+
+x = var(10)
+y = var(20)
+cond = var(false)
+ie = pd.create_ifelseop(inputs=[x], output_num=1)
+with ie.true_block():
+    x = ie.inputs(true, 0)
+    z = operator.add(x, y)
+    ie.set_output(true, 0, operator.softmax(z))
+with ie.false_block():
+    x = ie.inputs(false, 0)
+    z = layer.fc(x)
+    ie.set_output(true, 0, operator.softmax(z))
+out = b(cond)
+```
+
+In both examples, the left branch computes `softmax(x+y)` and the right branch computes `fc(x)`.
+
+A difference is that variables in the C++ program contain scalar values, whereas those in the PaddlePaddle programs are mini-batches of instances.  The `ie.input(true, 0)` invocation returns instances in the 0-th input, `x`, that corresponds to true values in `cond` as the local variable `x`, where `ie.input(false, 0)` returns instances corresponding to false values.
+
+### Blocks with `for` and `RNNOp`
+
+The following RNN model from the [RNN design doc](./rnn.md)
+
+```python
+x = sequence([10, 20, 30])
+m = var(0)
+W = tensor()
+U = tensor()
+
+rnn = create_rnn(inputs=[input])
+with rnn.stepnet() as net:
+  x = net.set_inputs(0)
+  h = net.add_memory(init=m)
+  fc_out = pd.matmul(W, x)
+  hidden_out = pd.matmul(U, h.pre(n=1))
+  sum = pd.add_two(fc_out, hidden_out)
+  act = pd.sigmoid(sum)
+  h.update(act)                       # update memory with act
+  net.set_outputs(0, act, hidden_out) # two outputs
+
+o1, o2 = rnn()
+print o1, o2
+```
+
+has its equivalent C++ program as follows
+
+```c++
+int* x = {10, 20, 30};
+int m = 0;
+int W = some_value();
+int U = some_other_value();
+
+int mem[sizeof(x) / sizeof(x[0]) + 1];
+int o1[sizeof(x) / sizeof(x[0]) + 1];
+int o2[sizeof(x) / sizeof(x[0]) + 1];
+for (int i = 1; i <= sizeof(x)/sizeof(x[0]); ++i) {
+  int x = x[i-1];
+  if (i == 1) mem[0] = m;
+  int fc_out = W * x;
+  int hidden_out = Y * mem[i-1];
+  int sum = fc_out + hidden_out;
+  int act = sigmoid(sum);
+  mem[i] = act;
+  o1[i] = act;
+  o2[i] = hidden_out;
+}
+
+print_array(o1);
+print_array(o2);
+```
+
+
+## Compilation and Execution
+
+Like TensorFlow programs, a PaddlePaddle program is written in Python.  The first part describes a neural network as a protobuf message, and the rest part executes the message for training or inference.
+
+The generation of this protobuf message is like what a compiler generates a binary executable file.  The execution of the message that the OS executes the binary file.
+
+## The "Binary Executable File Format"
+
+The definition of the protobuf message is as follows:
+
+```protobuf
+message BlockDesc {
+  repeated VarDesc vars = 1;
+  repeated OpDesc ops = 2;
+}
+```
+
+The step net in above RNN example would look like
+
+```
+BlockDesc {
+  vars = {
+    VarDesc {...} // x
+    VarDesc {...} // h
+    VarDesc {...} // fc_out
+    VarDesc {...} // hidden_out
+    VarDesc {...} // sum
+    VarDesc {...} // act
+  }
+  ops = {
+    OpDesc {...} // matmul
+    OpDesc {...} // add_two
+    OpDesc {...} // sigmoid
+  }
+};
+```
+
+Also, the RNN operator in above example is serialized into a protobuf message of type `OpDesc` and would look like:
+
+```
+OpDesc {
+  inputs = {0} // the index of x
+  outputs = {5, 3} // indices of act and hidden_out
+  attrs {
+    "memories" : {1} // the index of h
+    "step_net" : <above step net>
+  }
+};
+```
+
+This `OpDesc` value is in the `ops` field of the `BlockDesc` value representing the global block.
+
+
+## The Compilation of Blocks
+
+During the generation of the Protobuf message, the Block should store VarDesc (the Protobuf message which describes Variable) and OpDesc (the Protobuf message which describes Operator).
+
+VarDesc in a block should have its name scope to avoid local variables affect parent block's name scope.
+Child block's name scopes should inherit the parent's so that OpDesc in child block can reference a VarDesc that stored in parent block. For example
+
+```python
+a = pd.Varaible(shape=[20, 20])
+b = pd.fc(a, params=["fc.w", "fc.b"])
+
+rnn = pd.create_rnn()
+with rnn.stepnet() as net:
+    x = net.set_inputs(a)
+    # reuse fc's parameter
+    fc_without_b = pd.get_variable("fc.w")
+    net.set_outputs(fc_without_b)
+
+out = rnn()
+```
+the method `pd.get_variable` can help retrieve a Variable by a name, a Variable may store in a parent block, but might be retrieved in a child block, so block should have a variable scope that supports inheritance.
+
+In compiler design, the symbol table is a data structure created and maintained by compilers to store information about the occurrence of various entities such as variable names, function names, classes, etc.
+
+To store the definition of variables and operators, we define a C++ class `SymbolTable`, like the one used in compilers.
+
+`SymbolTable` can do the following stuff:
+
+- store the definitions (some names and attributes) of variables and operators,
+- to verify if a variable was declared,
+- to make it possible to implement type checking (offer Protobuf message pointers to `InferShape` handlers).
+
+
+```c++
+// Information in SymbolTable is enough to trace the dependency graph. So maybe
+// the Eval() interface takes a SymbolTable is enough.
+class SymbolTable {
+ public:
+  SymbolTable(SymbolTable* parent) : parent_(parent) {}
+
+  OpDesc* NewOp(const string& name="");
+
+  // TODO determine whether name is generated by python or C++
+  // currently assume that a unique name will be generated by C++ if the
+  // argument name left default.
+  VarDesc* NewVar(const string& name="");
+
+  // find a VarDesc by name, if recursive true, find parent's SymbolTable
+  // recursively.
+  // this interface is introduced to support InferShape, find protobuf messages
+  // of variables and operators, pass pointers into InferShape.
+  // operator
+  //
+  // NOTE maybe some C++ classes such as VarDescBuilder and OpDescBuilder should
+  // be proposed and embedded into pybind to enable python operate on C++ pointers.
+  VarDesc* FindVar(const string& name, bool recursive=true);
+
+  OpDesc* FindOp(const string& name);
+
+  BlockDesc Compile() const;
+
+ private:
+  SymbolTable* parent_;
+
+  map<string, OpDesc> ops_;
+  map<string, VarDesc> vars_;
+};
+```
+
+After all the description of variables and operators is added into SymbolTable,
+the block has enough information to run.
+
+The `Block` class takes a `BlockDesc` as input, and provide `Run` and `InferShape` functions.
+
+
+```c++
+namespace {
+
+class Block : OperatorBase {
+public:
+  Block(const BlockDesc& desc) desc_(desc) {}
+
+  void InferShape(const framework::Scope& scope) const override {
+    if (!symbols_ready_) {
+      CreateVariables(scope);
+      CreateOperators();
+    }
+    // should run InferShape first.
+    for (auto& op : runtime_table_.ops()) {
+      op->InferShape(scope);
+    }
+  }
+
+  void Run(const framework::Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override {
+    PADDLE_ENFORCE(symbols_ready_, "operators and variables should be created first.");
+    for (auto& op : runtime_table_.ops()) {
+      op->Run(scope, dev_ctx);
+    }
+  }
+
+  void CreateVariables(const framework::Scope& scope);
+  void CreateOperators();
+
+  // some other necessary interfaces of NetOp are list below
+  // ...
+
+private:
+  BlockDesc desc_;
+  bool symbols_ready_{false};
+};
+```
+
+## The Execution of Blocks
+
+Block inherits from OperatorBase, which has a Run method.
+Block's Run method will run its operators sequentially.
+
+There is another important interface called `Eval`, which take some arguments called targets, and generate a minimal graph which takes targets as the end points and creates a new Block,
+after `Run`, `Eval` will get the latest value and return the targets.
+
+The definition of Eval is as follows:
+
+```c++
+// clean a block description by targets using the corresponding dependency graph.
+// return a new BlockDesc with minimal number of operators.
+// NOTE not return a Block but the block's description so that this can be distributed
+// to a cluster.
+BlockDesc Prune(const BlockDesc& desc, vector<string> targets);
+
+void Block::Eval(const vector<string>& targets,
+                 const framework::Scope& scope,
+                 const platform::DeviceContext& dev_ctx) {
+  BlockDesc min_desc = Prune(desc_, targets);
+  Block min_block(min_desc);
+  min_block.Run(scope, dev_ctx);
+}
+```

doc/design/if_else_op.md

-IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has M (M<=N) instances, each corresponds to a true element in `cond`.
-
-```python
-import paddle as pd
-
-x = var()
-y = var()
-cond = var()
-
-b = pd.create_ifop(inputs=[x], output_num=1)
-with b.true_block():
-    x = b.inputs(0)
-    z = operator.add(x, y)
-    b.set_output(0, operator.softmax(z))
-
-out = b(cond)
-```
-
-If we want the output still has N instances, we can use IfElseOp with a default value, whose minibatch size must be N:
+IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has N instances. If cond[i] == True, input instance input[i] will go through true_block() and generate output[i]; otherwise it will produce output from false_bloack().
 
 ```python
 import paddle as pd
@@ -39,7 +21,7 @@ with b.false_block():
 out = b(cond)
 ```
 
-If only true_block is set in an IfElseOp, we can have a default value for false as:
+If only true_block is set in an IfElseOp, a special case is that we can have a default value for false as:
 ```python
 import paddle as pd
 

paddle/cuda/include/hl_tensor_ops.h

@@ -461,7 +461,7 @@ class add<float32x4_t> {
 public:
   INLINE float32x4_t operator()(const float32x4_t a,
                                 const float32x4_t b) const {
-    return vmulq_f32(a, b);
+    return vaddq_f32(a, b);
   }
 };
 

paddle/framework/tensor_impl.h

@@ -22,7 +22,7 @@ namespace framework {
 template <typename T>
 inline void Tensor::check_memory_size() const {
   PADDLE_ENFORCE_NOT_NULL(
-      holder_, "Tenosr holds no memory. Call Tensor::mutable_data first.");
+      holder_, "Tensor holds no memory. Call Tensor::mutable_data first.");
   PADDLE_ENFORCE_GE(
       holder_->size(), numel() * sizeof(T) + offset_,
       "Tensor's dims_ is out of bound. Call Tensor::mutable_data "

paddle/framework/tensor_test.cc

@@ -36,7 +36,7 @@ TEST(Tensor, DataAssert) {
   } catch (paddle::platform::EnforceNotMet err) {
     caught = true;
     std::string msg =
-        "holder_ should not be null\nTenosr holds no memory. Call "
+        "holder_ should not be null\nTensor holds no memory. Call "
         "Tensor::mutable_data first.";
     const char* what = err.what();
     for (size_t i = 0; i < msg.length(); ++i) {
@@ -112,7 +112,7 @@ TEST(Tensor, ShareDataWith) {
     } catch (paddle::platform::EnforceNotMet err) {
       caught = true;
       std::string msg =
-          "holder_ should not be null\nTenosr holds no memory. Call "
+          "holder_ should not be null\nTensor holds no memory. Call "
           "Tensor::mutable_data first.";
       const char* what = err.what();
       for (size_t i = 0; i < msg.length(); ++i) {

paddle/gserver/layers/ExpandConvBaseLayer.cpp

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#include "ExpandConvBaseLayer.h"
-
-#include "paddle/utils/Logging.h"
-namespace paddle {
-
-bool ExpandConvBaseLayer::init(const LayerMap &layerMap,
-                               const ParameterMap &parameterMap) {
-  /* Initialize the basic convolutional parent class */
-  ConvBaseLayer::init(layerMap, parameterMap);
-
-  int index = 0;
-  for (auto &inputConfig : config_.inputs()) {
-    const ConvConfig &conf = inputConfig.conv_conf();
-    /* Consistent caffe mode for multiple input */
-    caffeMode_ = conf.caffe_mode();
-
-    // create a new weight
-    size_t height, width;
-    height = filterPixels_[index] * filterChannels_[index];
-    width = (!isDeconv_) ? numFilters_ : channels_[index];
-    CHECK_EQ(parameters_[index]->getSize(), width * height);
-    Weight *w = new Weight(height, width, parameters_[index]);
-    weights_.emplace_back(w);
-    index++;
-  }
-  if (biasParameter_.get()) {
-    if (sharedBiases_) {
-      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
-      biases_ =
-          std::unique_ptr<Weight>(new Weight(numFilters_, 1, biasParameter_));
-    } else {
-      biases_ =
-          std::unique_ptr<Weight>(new Weight(getSize(), 1, biasParameter_));
-    }
-  }
-  getOutputSize();
-
-  return true;
-}
-
-size_t ExpandConvBaseLayer::getOutputSize() {
-  CHECK_NE(inputLayers_.size(), 0UL);
-  size_t layerSize = ConvBaseLayer::calOutputSize();
-  return layerSize;
-}
-
-void ExpandConvBaseLayer::addSharedBias() {
-  size_t mapW = getOutputSize() / numFilters_;
-  size_t mapH = getOutputValue()->getElementCnt() / mapW;
-  MatrixPtr out =
-      Matrix::create(getOutputValue()->getData(), mapH, mapW, false, useGpu_);
-
-  Matrix::resizeOrCreate(transOutValue_, mapW, mapH, false, useGpu_);
-
-  out->transpose(transOutValue_, false);  // false means no memory allocation
-  transOutValue_->reshape(transOutValue_->getElementCnt() / numFilters_,
-                          numFilters_);
-
-  MatrixPtr bias = Matrix::create(biases_->getW()->getData(),
-                                  1,
-                                  biases_->getW()->getElementCnt(),
-                                  false,
-                                  useGpu_);
-  transOutValue_->addBias(*bias, 1.0f);
-
-  transOutValue_->reshape(mapW, mapH);
-  transOutValue_->transpose(out, false);  // false means no memory allocation
-
-  out->clear();
-  bias->clear();
-}
-
-void ExpandConvBaseLayer::addUnsharedBias() {
-  MatrixPtr outValue = getOutputValue();
-  MatrixPtr bias = Matrix::create(biases_->getW()->getData(),
-                                  1,
-                                  biases_->getW()->getElementCnt(),
-                                  false,
-                                  useGpu_);
-  outValue->addBias(*bias, 1.0f);
-}
-
-void ExpandConvBaseLayer::bpropSharedBias(MatrixPtr biases, MatrixPtr v) {
-  size_t mapW = getOutputSize() / numFilters_;
-  size_t mapH = v->getElementCnt() / mapW;
-  MatrixPtr vTmp = Matrix::create(v->getData(), mapH, mapW, false, useGpu_);
-
-  Matrix::resizeOrCreate(transOutValue_, mapW, mapH, false, useGpu_);
-
-  vTmp->transpose(transOutValue_, false);  // false means no memory allocation
-  transOutValue_->reshape(transOutValue_->getElementCnt() / numFilters_,
-                          numFilters_);
-  biases->collectBias(*transOutValue_, 1.0f);
-}
-
-void ExpandConvBaseLayer::bpropBiases(MatrixPtr v) {
-  MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(),
-                                    1,
-                                    biases_->getWGrad()->getElementCnt(),
-                                    false,
-                                    useGpu_);
-  if (sharedBiases_) {
-    bpropSharedBias(biases, v);
-  } else {
-    biases->collectBias(*v, 1.0f);
-  }
-  biases->clear();
-}
-
-}  // namespace paddle

paddle/gserver/layers/ExpandConvLayer.cpp

@@ -36,7 +36,36 @@ inline bool isDepthwiseConv(int channels, int groups) {
 bool ExpandConvLayer::init(const LayerMap &layerMap,
                            const ParameterMap &parameterMap) {
   /* Initialize the basic convolutional parent class */
-  ExpandConvBaseLayer::init(layerMap, parameterMap);
+  ConvBaseLayer::init(layerMap, parameterMap);
+
+  int index = 0;
+  for (auto &inputConfig : config_.inputs()) {
+    const ConvConfig &conf = inputConfig.conv_conf();
+    /* Consistent caffe mode for multiple input */
+    caffeMode_ = conf.caffe_mode();
+
+    // create a new weight
+    size_t height, width;
+    height = filterPixels_[index] * filterChannels_[index];
+    width = (!isDeconv_) ? numFilters_ : channels_[index];
+    CHECK_EQ(parameters_[index]->getSize(), width * height);
+    Weight *w = new Weight(height, width, parameters_[index]);
+    weights_.emplace_back(w);
+    index++;
+  }
+
+  if (biasParameter_.get()) {
+    if (sharedBiases_) {
+      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
+      biases_ = std::unique_ptr<Weight>(
+          new Weight(1, numFilters_, biasParameter_, 0));
+    } else {
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_, 0));
+    }
+  }
+
+  getOutputSize();
 
   size_t numInputs = config_.inputs_size();
   inputShape_.resize(numInputs);
@@ -108,6 +137,12 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
   return true;
 }
 
+size_t ExpandConvLayer::getOutputSize() {
+  CHECK_NE(inputLayers_.size(), 0UL);
+  size_t layerSize = ConvBaseLayer::calOutputSize();
+  return layerSize;
+}
+
 // i is the index of input layers
 #define BACKWARD_INPUT(i, inputs, outputs) \
   backward_[2 * i]->calc(inputs, outputs)
@@ -155,11 +190,7 @@ void ExpandConvLayer::forward(PassType passType) {
 
   /* add the bias-vector */
   if (biases_.get()) {
-    if (sharedBiases_) {
-      addSharedBias();
-    } else {
-      addUnsharedBias();
-    }
+    output_.value->addBias(*biases_->getW(), 1.0, sharedBiases_);
   }
 
   /* activation */
@@ -171,7 +202,7 @@ void ExpandConvLayer::backward(const UpdateCallback &callback) {
 
   MatrixPtr outGrad = getOutputGrad();
   if (biases_ && biases_->getWGrad()) {
-    bpropBiases(outGrad);
+    biases_->getWGrad()->collectBias(*getOutputGrad(), 1, sharedBiases_);
     /* Increasing the number of gradient */
     biases_->getParameterPtr()->incUpdate(callback);
   }

paddle/gserver/layers/ExpandConvLayer.h

@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <vector>
-#include "ExpandConvBaseLayer.h"
+#include "ConvBaseLayer.h"
 #include "paddle/math/Matrix.h"
 
 namespace paddle {
@@ -28,10 +28,9 @@ namespace paddle {
  * The config file api is img_conv_layer.
  */
 
-class ExpandConvLayer : public ExpandConvBaseLayer {
+class ExpandConvLayer : public ConvBaseLayer {
 public:
-  explicit ExpandConvLayer(const LayerConfig& config)
-      : ExpandConvBaseLayer(config) {}
+  explicit ExpandConvLayer(const LayerConfig& config) : ConvBaseLayer(config) {}
 
   ~ExpandConvLayer() {}
 
@@ -41,6 +40,8 @@ public:
   void forward(PassType passType) override;
   void backward(const UpdateCallback& callback) override;
 
+  size_t getOutputSize();
+
 protected:
   std::vector<TensorShape> inputShape_;
   std::vector<TensorShape> filterShape_;

paddle/gserver/layers/MKLDNNConvLayer.cpp

@@ -285,10 +285,9 @@ void MKLDNNConvLayer::resetWgtBiasValue(
   wgt = MKLDNNMatrix::create(weight_->getW(), pd->weights_primitive_desc());
   VLOG(MKLDNN_FMTS) << "Weight value format: " << wgt->getFormat();
 
-  bias = nullptr;
-  if (biases_ && biases_->getW()) {
-    bias = MKLDNNMatrix::create(biases_->getW(), pd->bias_primitive_desc());
-  }
+  bias = (biases_ && biases_->getW())
+             ? MKLDNNMatrix::create(biases_->getW(), pd->bias_primitive_desc())
+             : nullptr;
 }
 
 void MKLDNNConvLayer::resetOutValue(
@@ -356,6 +355,7 @@ void MKLDNNConvLayer::resetBwdWgtPD(
 
 void MKLDNNConvLayer::resetBwdDataPD(
     std::shared_ptr<conv_bwdData::primitive_desc>& pd) {
+  pd = nullptr;
   if (inputLayers_[0]->getOutput().grad == nullptr) {
     return;
   }
@@ -476,6 +476,7 @@ void MKLDNNConvLayer::resetWgtBiasGrad(
       << "primitive desc of weight grad and value should be equal";
   VLOG(MKLDNN_FMTS) << "weight grad format: " << wgt->getFormat();
 
+  bias = nullptr;
   if (biasVal_ == nullptr) {
     return;
   }

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -17,9 +17,6 @@ limitations under the License. */
 
 using namespace mkldnn;  // NOLINT
 typedef memory::format format;
-typedef inner_product_forward fc_fwd;
-typedef inner_product_backward_weights fc_bwdWgt;
-typedef inner_product_backward_data fc_bwdData;
 
 namespace paddle {
 
@@ -93,35 +90,88 @@ void MKLDNNFcLayer::reshape(
   printSizeInfo();
 }
 
-void MKLDNNFcLayer::resetFwd(std::vector<mkldnn::primitive>& pipeline,
+void MKLDNNFcLayer::resetFwd(std::vector<primitive>& pipeline,
                              MKLDNNMatrixPtr& in,
                              MKLDNNMatrixPtr& wgt,
                              MKLDNNMatrixPtr& bias,
                              MKLDNNMatrixPtr& out) {
-  pipeline.clear();
-  bool hasBias = biases_ && biases_->getW();
-  const MatrixPtr& wgtVal = weight_->getW();
-  const MatrixPtr& biasVal = hasBias ? biases_->getW() : nullptr;
-  const MatrixPtr& outVal = output_.value;
+  resetFwdBuffers(in, wgt, bias, out);
+
+  resetFwdPD(fwdPD_, in, wgt, bias, out);
+
+  resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
+
+  printValueFormatFlow();
+}
+
+void MKLDNNFcLayer::resetBwd(std::vector<primitive>& pipeline,
+                             MKLDNNMatrixPtr& in,
+                             MKLDNNMatrixPtr& wgt,
+                             MKLDNNMatrixPtr& bias,
+                             MKLDNNMatrixPtr& out) {
+  std::shared_ptr<fc_bwdWgt::primitive_desc> bwdWgtPD;
+  std::shared_ptr<fc_bwdData::primitive_desc> bwdDataPD;
+
+  resetBwdBuffers(in, wgt, bias, out);
+
+  resetBwdWgtPD(bwdWgtPD, wgt, bias, out);
+
+  resetBwdDataPD(bwdDataPD, in, out);
 
+  resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out);
+
+  printGradFormatFlow();
+}
+
+void MKLDNNFcLayer::updateInputData() {
+  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+}
+
+void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
+  weight_->getParameterPtr()->incUpdate(callback);
+  if (biases_ && biases_->getWGrad()) {
+    biases_->getParameterPtr()->incUpdate(callback);
+  }
+}
+
+void MKLDNNFcLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
+                                    MKLDNNMatrixPtr& wgt,
+                                    MKLDNNMatrixPtr& bias,
+                                    MKLDNNMatrixPtr& out) {
+  resetInValue(in);
+
+  resetWgtBiasValue(wgt, bias);
+
+  resetOutValue(out);
+}
+
+void MKLDNNFcLayer::resetInValue(MKLDNNMatrixPtr& in) {
   if (inputIsOnlyMKLDNN()) {
-    const MatrixPtr& inVal = getInputValue(0);
-    in = std::dynamic_pointer_cast<MKLDNNMatrix>(inVal);
+    const MatrixPtr& dnnIn = getInputValue(0);
+    in = std::dynamic_pointer_cast<MKLDNNMatrix>(dnnIn);
     CHECK(in) << "Input should be MKLDNNMatrix";
   } else {
     CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
-    const MatrixPtr& inVal = getInputValue(0, CPU_DEVICE);
+    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
     in = MKLDNNMatrix::create(
-        inVal, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_);
+        cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
   }
   in->downSpatial();
+}
+
+void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt,
+                                      MKLDNNMatrixPtr& bias) {
   wgt = MKLDNNMatrix::create(
-      wgtVal, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_);
+      weight_->getW(), {oc_, ic_, ih_, iw_}, format::oihw, engine_);
   wgt->downSpatial();
-  bias = hasBias ? MKLDNNMatrix::create(biasVal, {oc_}, format::x, engine_)
+
+  bias = (biases_ && biases_->getW())
+             ? MKLDNNMatrix::create(biases_->getW(), {oc_}, format::x, engine_)
              : nullptr;
-  out = MKLDNNMatrix::create(outVal, {bs_, oc_}, format::nc, engine_);
+}
 
+void MKLDNNFcLayer::resetOutValue(MKLDNNMatrixPtr& out) {
+  out = MKLDNNMatrix::create(output_.value, {bs_, oc_}, format::nc, engine_);
   // change original output value to mkldnn output value
   output_.value = std::dynamic_pointer_cast<Matrix>(out);
   if (!outputIsOnlyMKLDNN()) {
@@ -129,10 +179,18 @@ void MKLDNNFcLayer::resetFwd(std::vector<mkldnn::primitive>& pipeline,
     // just share point
     getOutput(CPU_DEVICE).value->setData(output_.value->getData());
   }
+}
 
-  // create forward handle
+void MKLDNNFcLayer::resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
+                               MKLDNNMatrixPtr in,
+                               MKLDNNMatrixPtr wgt,
+                               MKLDNNMatrixPtr bias,
+                               MKLDNNMatrixPtr out) {
+  CHECK(in);
+  CHECK(wgt);
+  CHECK(out);
   prop_kind pk = prop_kind::forward;
-  fc_fwd::desc fwdDesc = hasBias ? fc_fwd::desc(pk,
+  fc_fwd::desc fwdDesc = bias != nullptr ? fc_fwd::desc(pk,
                                                         in->getMemoryDesc(),
                                                         wgt->getMemoryDesc(),
                                                         bias->getMemoryDesc(),
@@ -141,34 +199,39 @@ void MKLDNNFcLayer::resetFwd(std::vector<mkldnn::primitive>& pipeline,
                                                         in->getMemoryDesc(),
                                                         wgt->getMemoryDesc(),
                                                         out->getMemoryDesc());
-  fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-  if (hasBias) {
-    fwd_.reset(new fc_fwd(fwdPD, *in, *wgt, *bias, *out));
+  pd.reset(new fc_fwd::primitive_desc(fwdDesc, engine_));
+}
+
+void MKLDNNFcLayer::resetFwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<fc_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+
+  if (bias) {
+    fwd_.reset(new fc_fwd(*pd, *in, *wgt, *bias, *out));
   } else {
-    fwd_.reset(new fc_fwd(fwdPD, *in, *wgt, *out));
+    fwd_.reset(new fc_fwd(*pd, *in, *wgt, *out));
   }
-  printValueFormatFlow();
 
   pipeline.push_back(*fwd_);
 }
 
-void MKLDNNFcLayer::resetBwd(std::vector<mkldnn::primitive>& pipeline,
-                             MKLDNNMatrixPtr& in,
+void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
                                     MKLDNNMatrixPtr& wgt,
                                     MKLDNNMatrixPtr& bias,
                                     MKLDNNMatrixPtr& out) {
-  pipeline.clear();
-  if (!needResetBwd_) {
-    return;
-  }
-  needResetBwd_ = false;
-  bool hasBias = biases_ && biases_->getWGrad();
+  resetOutGrad(out);
 
-  /// backward weight
-  CHECK(inVal_) << "Should have input value";
-  const MatrixPtr& wgtGrad = weight_->getWGrad();
-  const MatrixPtr& biasGrad = hasBias ? biases_->getWGrad() : nullptr;
+  resetWgtBiasGrad(wgt, bias);
 
+  resetInGrad(in);
+}
+
+void MKLDNNFcLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
   // TODO(TJ): merge outgrad
   int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
   // for MKLDNN device:
@@ -178,66 +241,88 @@ void MKLDNNFcLayer::resetBwd(std::vector<mkldnn::primitive>& pipeline,
   // for CPU device:
   // fc do not need to convert from cpu device since output is always nc format
   // only need create from cpu device
-  const MatrixPtr& outGrad = getOutput(device).grad;
-  out = MKLDNNMatrix::create(outGrad, outVal_->getPrimitiveDesc());
-  wgt = MKLDNNMatrix::create(wgtGrad, wgtVal_->getPrimitiveDesc());
-  bias = hasBias ? MKLDNNMatrix::create(biasGrad, biasVal_->getPrimitiveDesc())
-                 : nullptr;
+  CHECK(outVal_);
+  out =
+      MKLDNNMatrix::create(getOutput(device).grad, outVal_->getPrimitiveDesc());
+}
 
-  // create memory primitive desc
-  fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward,
-                                      inVal_->getMemoryDesc(),
-                                      wgt->getMemoryDesc(),
-                                      out->getMemoryDesc());
-  fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-  fc_bwdWgt::desc bwdWgtDesc = hasBias
-                                   ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
+void MKLDNNFcLayer::resetWgtBiasGrad(MKLDNNMatrixPtr& wgt,
+                                     MKLDNNMatrixPtr& bias) {
+  CHECK(wgtVal_);
+  wgt = MKLDNNMatrix::create(weight_->getWGrad(), wgtVal_->getPrimitiveDesc());
+
+  bias = nullptr;
+  if (biasVal_ == nullptr) {
+    return;
+  }
+  bias =
+      MKLDNNMatrix::create(biases_->getWGrad(), biasVal_->getPrimitiveDesc());
+}
+
+void MKLDNNFcLayer::resetInGrad(MKLDNNMatrixPtr& in) {
+  in = nullptr;
+  const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
+  if (inGrad == nullptr) {
+    return;
+  }
+  // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
+  CHECK(inVal_);
+  in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
+}
+
+void MKLDNNFcLayer::resetBwdWgtPD(
+    std::shared_ptr<fc_bwdWgt::primitive_desc>& pd,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  CHECK(inVal_);
+  fc_bwdWgt::desc bwdWgtDesc = bias ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
                                                       wgt->getMemoryDesc(),
                                                       bias->getMemoryDesc(),
                                                       out->getMemoryDesc())
                                     : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
                                                       wgt->getMemoryDesc(),
                                                       out->getMemoryDesc());
-  fc_bwdWgt::primitive_desc bwdWgtPD =
-      fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD);
+  pd.reset(new fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_));
+}
+
+void MKLDNNFcLayer::resetBwdDataPD(
+    std::shared_ptr<fc_bwdData::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& out) {
+  pd = nullptr;
+  if (in == nullptr) {
+    return;
+  }
+  CHECK(wgtVal_);
+  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(
+      in->getMemoryDesc(), wgtVal_->getMemoryDesc(), out->getMemoryDesc());
+  pd.reset(new fc_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_));
+}
 
-  if (hasBias) {
-    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt, *bias));
+void MKLDNNFcLayer::resetBwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<fc_bwdWgt::primitive_desc>& bwdWgtPD,
+    std::shared_ptr<fc_bwdData::primitive_desc>& bwdDataPD,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+  CHECK(inVal_);
+  if (bias) {
+    bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVal_, *out, *wgt, *bias));
   } else {
-    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt));
+    bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVal_, *out, *wgt));
   }
   pipeline.push_back(*bwdWgt_);
 
-  /// backward data
-  const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
-  if (inGrad == nullptr) {
+  if (bwdDataPD == nullptr) {
     return;
   }
-  if (getInput(0, MKLDNN_DEVICE).getAllCount() > 1) {
-    // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
-  } else {
-    in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
-  }
-
-  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(
-      inVal_->getMemoryDesc(), wgt->getMemoryDesc(), out->getMemoryDesc());
-  fc_bwdData::primitive_desc bwdDataPD =
-      fc_bwdData::primitive_desc(bwdDataDesc, engine_, fwdPD);
-
   CHECK(wgtVal_) << "Should have weight memory";
-  bwdData_.reset(new fc_bwdData(bwdDataPD, *out, *wgtVal_, *in));
-  printGradFormatFlow();
+  bwdData_.reset(new fc_bwdData(*bwdDataPD, *out, *wgtVal_, *in));
   pipeline.push_back(*bwdData_);
 }
 
-void MKLDNNFcLayer::updateInputData() {
-  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
-}
-
-void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
-  weight_->getParameterPtr()->incUpdate(callback);
-  if (biases_ && biases_->getWGrad()) {
-    biases_->getParameterPtr()->incUpdate(callback);
-  }
-}
 }  // namespace paddle

paddle/gserver/layers/MKLDNNFcLayer.h

@@ -18,6 +18,9 @@ limitations under the License. */
 #include "mkldnn.hpp"
 
 namespace paddle {
+typedef mkldnn::inner_product_forward fc_fwd;
+typedef mkldnn::inner_product_backward_weights fc_bwdWgt;
+typedef mkldnn::inner_product_backward_data fc_bwdData;
 
 /**
  * @brief A subclass of MKLDNNLayer fc layer.
@@ -32,6 +35,9 @@ protected:
   // if has already init the weight
   bool hasInitedWgt_;
 
+  // save forward primitive_desc, which can be used backward
+  std::shared_ptr<fc_fwd::primitive_desc> fwdPD_;
+
   // fc weight and bias
   std::unique_ptr<Weight> weight_;
   std::unique_ptr<Weight> biases_;
@@ -67,6 +73,59 @@ public:
   void convertWeightsFromPaddle() override;
 
   void convertWeightsToPaddle() override;
+
+protected:
+  /**
+   * Forward functions: reset buffers(input, output, weight and bias),
+   *                    reset primitive descriptor,
+   *                    reset pipeline.
+   */
+  void resetFwdBuffers(MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  void resetInValue(MKLDNNMatrixPtr& in);
+  void resetWgtBiasValue(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
+  void resetOutValue(MKLDNNMatrixPtr& out);
+  void resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
+                  MKLDNNMatrixPtr in,
+                  MKLDNNMatrixPtr wgt,
+                  MKLDNNMatrixPtr bias,
+                  MKLDNNMatrixPtr out);
+  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<fc_fwd::primitive_desc>& pd,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * Backward functions: reset buffers(input, output, weight and bias),
+   *                     reset primitive descriptor for backward weight,
+   *                     reset primitive descriptor for backward data,
+   *                     reset pipeline.
+   */
+  void resetBwdBuffers(MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  void resetOutGrad(MKLDNNMatrixPtr& out);
+  void resetWgtBiasGrad(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
+  void resetInGrad(MKLDNNMatrixPtr& in);
+  void resetBwdWgtPD(std::shared_ptr<fc_bwdWgt::primitive_desc>& pd,
+                     MKLDNNMatrixPtr& wgt,
+                     MKLDNNMatrixPtr& bias,
+                     MKLDNNMatrixPtr& out);
+  void resetBwdDataPD(std::shared_ptr<fc_bwdData::primitive_desc>& pd,
+                      MKLDNNMatrixPtr& in,
+                      MKLDNNMatrixPtr& out);
+  void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<fc_bwdWgt::primitive_desc>& bwdWgtPD,
+                        std::shared_ptr<fc_bwdData::primitive_desc>& bwdDataPD,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
 };
 
 }  // namespace paddle

paddle/math/BaseMatrix.cu

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <cmath>
 #include "BaseMatrix.h"
 #include "MathFunctions.h"
+#include "NEONFunctions.h"
 #include "SIMDFunctions.h"
 #include "hl_matrix_apply.cuh"
 #include "hl_matrix_base.cuh"
@@ -666,6 +667,13 @@ void BaseMatrixT<T>::relu(BaseMatrixT& b) {
   applyBinary(binary::Relu<T>(), b);
 }
 
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+template <>
+void BaseMatrixT<float>::relu(BaseMatrixT& b) {
+  neon::relu(data_, b.data_, height_ * width_);
+}
+#endif
+
 DEFINE_MATRIX_BINARY_OP(ReluDerivative, a *= (b > 0.0f ? 1.0f : 0.0f));
 template <class T>
 void BaseMatrixT<T>::reluDerivative(BaseMatrixT& b) {

paddle/math/MKLDNNMatrix.h

@@ -66,11 +66,12 @@ public:
   /**
    * Create reorder primitive.
    * Create a mkldnn::reorder handle for converting src MKLDNNMatrix to dst.
-   * checkData: for whether to check the data handle of src and dst is the same.
-   *            if true, means check it and do not want support inplace reorder;
-   *            otherwise do not check data which means the created reorder
-   *            maybe inplace buffer and do not guarantee the logical is correct
-   *            since not all format or conversion support inplace.
+   * checkData: whether to check the data handle of src and dst.
+   *            if true, it will check the data and do not allow them equal;
+   *            otherwise, it will not check them, then the reorder created
+   *            may have inplace buffer.
+   *            Do not set false, if you can not guarantee the inplace logical
+   *            would work with your reorder.
    */
   static std::shared_ptr<mkldnn::reorder> createReorder(
       const MKLDNNMatrixPtr& src,

paddle/math/NEONFunctions.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+
+#include "NEONFunctions.h"
+#include <arm_neon.h>
+
+namespace paddle {
+namespace neon {
+
+// b[i] = a[i] > 0.0f ? a[i] : 0.0f
+void relu(const float* a, float* b, int len) {
+  int offset = len % 16;
+  float32x4_t ma0, ma1, ma2, ma3;
+  float32x4_t mb0, mb1, mb2, mb3;
+
+  float32x4_t zero = vdupq_n_f32(0.f);
+  for (int k = 0; k < len / 16; k++, a += 16, b += 16) {
+    ma0 = vld1q_f32(a);
+    ma1 = vld1q_f32(a + 4);
+    ma2 = vld1q_f32(a + 8);
+    ma3 = vld1q_f32(a + 12);
+
+    mb0 = vmaxq_f32(ma0, zero);
+    mb1 = vmaxq_f32(ma1, zero);
+    mb2 = vmaxq_f32(ma2, zero);
+    mb3 = vmaxq_f32(ma3, zero);
+
+    vst1q_f32(b, mb0);
+    vst1q_f32(b + 4, mb1);
+    vst1q_f32(b + 8, mb2);
+    vst1q_f32(b + 12, mb3);
+  }
+
+  for (int i = 0; i < offset; i++) {
+    b[i] = a[i] > 0.0f ? a[i] : 0.0f;
+  }
+}
+
+}  // namespace neon
+}  // namespace paddle
+
+#endif

paddle/gserver/layers/ExpandConvBaseLayer.h → paddle/math/NEONFunctions.h

@@ -14,44 +14,10 @@ limitations under the License. */
 
 #pragma once
 
-#include <vector>
-#include "ConvBaseLayer.h"
-#include "paddle/math/Matrix.h"
-
 namespace paddle {
+namespace neon {
 
-/**
- * @brief A subclass of ConvBaseLayer that is a superclass of both
- * ExpandConvLayer and ExpandConvTransLayer
- */
-class ExpandConvBaseLayer : public ConvBaseLayer {
-protected:
-  /// The transpose of output, which is an auxiliary matrix.
-  MatrixPtr transOutValue_;
-
-public:
-  explicit ExpandConvBaseLayer(const LayerConfig& config)
-      : ConvBaseLayer(config) {}
-
-  ~ExpandConvBaseLayer() {}
-
-  bool init(const LayerMap& layerMap,
-            const ParameterMap& parameterMap) override;
-
-  size_t getOutputSize();
-
-  /**
-   * Add shared bias.
-   */
-  void addSharedBias();
-
-  /**
-   * Add unshared bias.
-   */
-  void addUnsharedBias();
-
-  void bpropSharedBias(MatrixPtr biases, MatrixPtr v);
-  void bpropBiases(MatrixPtr v);
-};
+void relu(const float* a, float* b, int len);
 
+}  // namespace neon
 }  // namespace paddle

paddle/memory/memcpy.cc

@@ -62,6 +62,24 @@ void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
   }
 }
 
+template <>
+void Copy<platform::CPUPlace, platform::GPUPlace>(platform::CPUPlace dst_place,
+                                                  void* dst,
+                                                  platform::GPUPlace src_place,
+                                                  const void* src, size_t num) {
+  platform::SetDeviceId(src_place.device);
+  platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToHost);
+}
+
+template <>
+void Copy<platform::GPUPlace, platform::CPUPlace>(platform::GPUPlace dst_place,
+                                                  void* dst,
+                                                  platform::CPUPlace src_place,
+                                                  const void* src, size_t num) {
+  platform::SetDeviceId(dst_place.device);
+  platform::GpuMemcpySync(dst, src, num, cudaMemcpyHostToDevice);
+}
+
 #endif  // PADDLE_ONLY_CPU
 
 }  // namespace memory

paddle/operators/CMakeLists.txt

@@ -80,9 +80,11 @@ endfunction()
 add_subdirectory(math)
 
 set(DEPS_OPS
-    recurrent_op)
+    recurrent_op
+    cond_op)
 op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
   DEPS framework_proto tensor net_op)
+op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op)
 
 list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
 foreach(src ${GENERAL_OPS})

paddle/operators/accuracy_op.cc

@@ -23,10 +23,15 @@ class AccuracyOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Inference"),
-                            "Input of Inference must be initialized.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("Inference"),
+        "Input(Inference) of AccuracyOp should not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
-                            "Input of Inference must be initialized.");
+                            "Input(Label) of AccuracyOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Accuracy"),
+        "Output(Accuracy) of AccuracyOp should not be null.");
+
     auto *inference = ctx.Input<framework::Tensor>("Inference");
     auto *label = ctx.Input<framework::Tensor>("Label");
 

paddle/operators/accuracy_op.cu

@@ -12,26 +12,38 @@ 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 <thrust/execution_policy.h>
+#include <thrust/reduce.h>
 #include "paddle/operators/accuracy_op.h"
+#include "paddle/platform/cuda_helper.h"
 
 namespace paddle {
 namespace operators {
+using platform::PADDLE_CUDA_NUM_THREADS;
 
-__global__ void AccuracySingleKernel(const int N, const int D, const int top_k,
-                                     const int* Xdata, const int* labelData,
-                                     float* accuracy) {
-  int correct = 0;
-  for (int row = 0; row < N; row++) {
-    const int label = labelData[row];
-    for (int col = 0; col < D; col++) {
-      const int pred = Xdata[row * D + col];
-      if (pred == label) {
-        ++correct;
+template <int BlockSize>
+__global__ void AccuracyCudaKernel(const int N, const int D, const int* Xdata,
+                                   const int* labeldata, float* accuracy) {
+  int count = 0;
+  __shared__ int total[BlockSize];
+
+  // support only 1 block
+  for (int i = threadIdx.x; i < (N); i += BlockSize) {
+    for (int j = 0; j < D; ++j) {
+      if (Xdata[i * D + j] == labeldata[i]) {
+        ++count;
         break;
       }
     }
   }
-  *accuracy = static_cast<float>(correct) / static_cast<float>(N);
+  total[threadIdx.x] = count;
+  __syncthreads();
+
+  // reduce the count with init value 0, and output accuracy.
+  int result = thrust::reduce(thrust::device, total, total + BlockSize, 0);
+  if (threadIdx.x == 0) {
+    *accuracy = static_cast<float>(result) / static_cast<float>(N);
+  }
 }
 
 template <typename T>
@@ -57,8 +69,8 @@ class AccuracyOpCUDAKernel : public framework::OpKernel {
       return;
     }
 
-    AccuracySingleKernel<<<1, 1>>>(num_samples, infer_width, 1, inference_data,
-                                   label_data, accuracy_data);
+    AccuracyCudaKernel<PADDLE_CUDA_NUM_THREADS><<<1, PADDLE_CUDA_NUM_THREADS>>>(
+        num_samples, infer_width, inference_data, label_data, accuracy_data);
   }
 };
 

paddle/operators/add_op.cc

@@ -23,6 +23,13 @@ class AddOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of AddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of AddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of AddOp should not be null.");
+
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
                       ctx.Input<Tensor>("Y")->dims(),
                       "Two input of Add Op's dimension must be same.");

paddle/operators/concat_op.cc

@@ -25,6 +25,9 @@ class ConcatOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ConcatOp should not be null.");
+
     auto ins = ctx.MultiInput<framework::Tensor>("X");
     auto *out = ctx.Output<framework::LoDTensor>("Out");
     size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));

paddle/operators/cond_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    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/operators/cond_op.h"
+
+#include <cstring>
+#include <sstream>
+
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/gather.h"
+#include "paddle/operators/net_op.h"
+#include "paddle/operators/scatter.h"
+
+namespace paddle {
+namespace operators {
+
+using Scope = framework::Scope;
+using Variable = framework::Variable;
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+using DDim = framework::DDim;
+
+void CondOp::CreateScope(const Scope& scope) const {
+  auto sub_scopes_var = scope.FindVar("SubScopes");
+  PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
+                          "Output(SubScopes) of CondOp should not be null.");
+  auto sub_scopes = sub_scopes_var->GetMutable<std::vector<Scope*>>();
+  auto& sub_scope = scope.NewScope();
+  sub_scopes->push_back(&sub_scope);
+}
+
+void CondOp::CreateIndexTensor(const Scope& scope) const {
+  auto index_tensors_var = scope.FindVar("IndexTensors");
+  PADDLE_ENFORCE_NOT_NULL(index_tensors_var,
+                          "Output(IndexTensors) of CondOp should not be null.");
+  auto& index_tensors =
+      *index_tensors_var->GetMutable<std::vector<LoDTensor>>();
+  index_tensors.push_back(LoDTensor());
+}
+
+void CondOp::InferShape(const Scope& scope) const {
+  auto sub_scopes_var = scope.FindVar("SubScopes");
+  PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
+                          "Output(SubScopes) of CondOp should not be null.");
+  auto& sub_scopes = *sub_scopes_var->GetMutable<std::vector<Scope*>>();
+
+  for (int i = 0; i < 2; ++i) {
+    // Create two sub scopes for true and false branches
+    // sub_scopes[0] for the true branch and sub_scopes[1] for the false
+    // branch
+    CreateScope(scope);
+
+    // Create two tensors for true and false indices
+    // index_tensors[0] for the true branch and index_tensors[1] for the false
+    // branch
+    CreateIndexTensor(scope);
+
+    PADDLE_ENFORCE(!Inputs("Xs").empty(),
+                   "Inputs(Xs) of CondOp can't be empty.");
+    for (auto& input : Inputs("Xs")) {
+      // Create a new tensor in sub-scope for input-type tensor
+      Variable* v = sub_scopes[i]->NewVar(input);
+      LoDTensor* sub_input = v->GetMutable<LoDTensor>();
+      sub_input->Resize(scope.FindVar(input)->GetMutable<LoDTensor>()->dims());
+    }
+
+    for (auto& output : (*sub_net_op_[i]).Outputs()) {
+      for (auto& var_name : output.second) {
+        sub_scopes[i]->NewVar(var_name);
+      }
+    }
+
+    // each net calls InferShape
+    sub_net_op_[i]->InferShape(*sub_scopes[i]);
+  }
+
+  for (auto& output : Outputs("Outs")) {
+    LoDTensor* tensor_t_out =
+        sub_scopes[0]->FindVar(output)->GetMutable<LoDTensor>();
+    PADDLE_ENFORCE_NOT_NULL(tensor_t_out, "True output should not be NULL");
+    LoDTensor* tensor_f_out =
+        sub_scopes[1]->FindVar(output)->GetMutable<LoDTensor>();
+    PADDLE_ENFORCE_NOT_NULL(tensor_f_out, "False output should not be NULL");
+
+    auto* tensor_out_var = scope.FindVar(output);
+    PADDLE_ENFORCE_NOT_NULL(tensor_out_var, "Output not found");
+    LoDTensor* tensor_out = tensor_out_var->GetMutable<LoDTensor>();
+    PADDLE_ENFORCE_NOT_NULL(tensor_t_out,
+                            "True output tensor should not be NULL");
+
+    // check output size should be same
+    PADDLE_ENFORCE_EQ(tensor_t_out->dims(), tensor_f_out->dims(),
+                      "Outputs not of the same shape");
+    tensor_out->Resize(tensor_t_out->dims());
+    // tensor_out->mutable_data<float>(tensor_out->dims(),
+    // platform::CPUPlace());
+    tensor_out->mutable_data<float>(platform::CPUPlace());
+  }
+}
+
+void CondOp::Run(const Scope& scope,
+                 const platform::DeviceContext& dev_ctx) const {
+  auto* sub_scopes_var = scope.FindVar("SubScopes");
+  PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
+                          "Output(SubScopes) of CondOp should not be null.");
+  auto sub_scopes = sub_scopes_var->Get<std::vector<Scope*>>();
+  auto* index_tensors_var = scope.FindVar("IndexTensors");
+  PADDLE_ENFORCE_NOT_NULL(index_tensors_var,
+                          "Output(IndexTensors) of CondOp should not be null.");
+  auto index_tensors = index_tensors_var->Get<std::vector<LoDTensor>>();
+
+  std::string cond_name = Input("Cond");
+  Variable* cond_var = scope.FindVar(cond_name);
+  PADDLE_ENFORCE_NOT_NULL(cond_var,
+                          "Input(Cond) of CondOp should not be null.");
+  const LoDTensor* cond = cond_var->GetMutable<LoDTensor>();
+
+  // Step 1: get the true/false index at runtime
+  // index_[0]: vector<int>, contains all index for cond[i] == true
+  // index_[1]: vector<int>, contains all index for cond[i] == false
+  for (int i = 0; i < 2; ++i) index_[i].clear();
+
+  const int* cond_data = cond->data<int>();
+  for (int i = 0; i < cond->dims()[0]; ++i) {
+    if (cond_data[i])
+      index_[0].push_back(i);
+    else
+      index_[1].push_back(i);
+  }
+
+  // put index_[0] and index_[1] into two tensors:
+  // index_tensor_[0] and index_tensor_[1]
+  DDim dim = paddle::framework::make_ddim({0});
+  for (int i = 0; i < 2; ++i) {
+    dim[0] = index_[i].size();
+    int* tmp_ptr =
+        index_tensors[i].mutable_data<int>(dim, platform::CPUPlace());
+    index_tensors[i].Resize(dim);
+    memcpy(tmp_ptr, index_[i].data(), dim[0] * sizeof(int));
+  }
+
+  // Step 2: collect data by calling gather
+  for (int i = 0; i < 2; ++i) {
+    // i= 0/i for True and False branches respectively
+    for (auto& input : Inputs("Xs")) {
+      // find Tensor
+      Variable* v = scope.FindVar(input);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_parent = v->GetMutable<LoDTensor>();
+
+      v = sub_scopes[i]->FindVar(input);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_child = v->GetMutable<LoDTensor>();
+
+      // Resize child
+      DDim dim = tensor_child->dims();
+      dim[0] = index_[i].size();
+      tensor_child->Resize(dim);
+      tensor_child->mutable_data<float>(dim, platform::CPUPlace());
+
+      Gather<float>(dev_ctx.GetPlace(), tensor_parent, &index_tensors[i],
+                    tensor_child);
+    }
+  }
+
+  // Step 3: run
+  for (int i = 0; i < 2; ++i) {
+    sub_net_op_[i]->Run(*sub_scopes[i], dev_ctx);
+  }
+
+  // Step 4: merge output results
+  PADDLE_ENFORCE(!Outputs("Outs").empty(),
+                 "Outputs(Outs) of CondOp can't be empty.");
+  for (int i = 0; i < 2; ++i) {
+    // i= 0/i for True and False branches respectively
+    for (auto& output : Outputs("Outs")) {
+      // find Tensor
+      Variable* v = scope.FindVar(output);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_parent = v->GetMutable<LoDTensor>();
+
+      v = sub_scopes[i]->FindVar(output);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_child = v->GetMutable<LoDTensor>();
+
+      ScatterUpdate<float>(dev_ctx.GetPlace(), tensor_child, &index_tensors[i],
+                           tensor_parent);
+    }
+  }
+}
+
+class CondOpProtoAndCheckerMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  CondOpProtoAndCheckerMaker(framework::OpProto* proto,
+                             framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Cond", "The condition, which is a bool vector");
+    AddInput("Xs", "Inputs of Subnets").AsDuplicable();
+    AddOutput("Outs", "Outputs of Cond_Op after merge").AsDuplicable();
+
+    AddOutput("SubScopes", "sub scopes for true and false branches");
+    AddOutput("IndexTensors", "Index Tensors contains indices for true/false");
+
+    AddComment(R"DOC(
+Sample dependent Cond Operator:
+Given Cond[i] as a 1/0 vector to indicate true/false
+The equation is: 
+Out[i] = subnet_t[i], if Cond[i] == true
+Out[i] = subnet_t[i], if Cond[i] == false
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_WITHOUT_GRADIENT(cond, paddle::operators::CondOp,
+                             paddle::operators::CondOpProtoAndCheckerMaker);

paddle/operators/cond_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    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 <vector>
+#include "glog/logging.h"
+#include "paddle/framework/ddim.h"
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/operator.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+
+/*
+ * @brief CondOp is a dynamic if-else Operator
+ *
+ * It has a input tensor named cond indicating which netop each instance will
+ * run.
+ *
+ * if cond == 1, it will run true_net, which is a NetOp.
+ *
+ * if cond == 0, it will run false_net, which is another NetOp.
+ */
+class CondOp : public framework::OperatorBase {
+ public:
+  CondOp(const std::string& type, const framework::VariableNameMap& inputs,
+         const framework::VariableNameMap& outputs,
+         const framework::AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {
+    index_.resize(2);
+    sub_net_op_.resize(2);
+  }
+
+  CondOp(const CondOp& o)
+      : framework::OperatorBase(
+            static_cast<const framework::OperatorBase&>(o)) {
+    // TODO(yuyang18): Implement copy ctor well.
+    PADDLE_THROW("Not implemented");
+  }
+
+  void CreateScope(const framework::Scope& scope) const;
+
+  void CreateIndexTensor(const framework::Scope& scope) const;
+
+  /*
+   * InferShape must be called before Run.
+   */
+  void InferShape(const framework::Scope& scope) const override;
+
+  /*
+   * Set True Block
+   */
+  void set_truenet(std::unique_ptr<OperatorBase>&& net) {
+    sub_net_op_[0] = std::move(net);
+  }
+
+  /*
+   * Set False Block
+   */
+  void set_falsenet(std::unique_ptr<OperatorBase>&& net) {
+    sub_net_op_[1] = std::move(net);
+  }
+
+  void Run(const framework::Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override;
+
+ private:
+  // sub_net_op_[0]: subnet_t
+  // sub_net_op_[1]: subnet_f
+  std::vector<std::unique_ptr<framework::OperatorBase>> sub_net_op_;
+
+  // index_[0]: True_index;
+  // index_[1]: False_index;
+  mutable std::vector<std::vector<int>> index_;
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/cos_sim_op.cc

@@ -26,8 +26,16 @@ class CosSimOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     // notnull check
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("XNorm"),
+                            "Output(XNorm) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("YNorm"),
+                            "Output(YNorm) of CosSimOp should not be null.");
 
     // shape check
     auto x_dims = ctx.Input<Tensor>("X")->dims();

paddle/operators/cos_sim_op.h

@@ -56,7 +56,7 @@ class CosSimKernel : public framework::OpKernel {
     x_norm.device(place) = x.square().sum(row_along).sqrt();
     y_norm.device(place) = y.square().sum(row_along).sqrt();
     if (rows_x == rows_y) {
-      auto xy = (x * y).sum(Eigen::array<int, 1>({1}));
+      auto xy = (x * y).sum(Eigen::array<int, 1>({{1}}));
       z.device(place) = xy / x_norm / y_norm;
     } else {
       Eigen::DSizes<int, 2> bcast(rows_x, 1);
@@ -134,7 +134,7 @@ class CosSimGradKernel : public framework::OpKernel {
         out_grad_y->mutable_data<T>(context.GetPlace());
         auto dy = EigenMatrix<T>::Reshape(*out_grad_y, 1);
         auto grad = x / norm_prod_bcast - z_bcast * y_bcast / y_snorm_bcast;
-        dy.device(place) = (dz_bcast * grad).sum(Eigen::array<int, 1>({0}));
+        dy.device(place) = (dz_bcast * grad).sum(Eigen::array<int, 1>({{0}}));
       }
     }
   }

paddle/operators/elementwise_mul_op.cc

@@ -25,8 +25,14 @@ class ElementWiseMulOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of ElementWiseMulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of ElementWiseMulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of ElementWiseMulOp should not be null.");
+
     auto x_dim = ctx.Input<Tensor>("X")->dims();
     auto y_dim = ctx.Input<Tensor>("Y")->dims();
     PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),

paddle/operators/elementwise_mul_op.h

@@ -13,10 +13,8 @@
    limitations under the License. */
 
 #pragma once
-#include <iostream>
 #include "paddle/framework/eigen.h"
 #include "paddle/framework/op_registry.h"
-#include "paddle/operators/math/math_function.h"
 
 namespace paddle {
 namespace operators {

paddle/operators/fc_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+
+class FCOp : public NetOp {
+ public:
+  FCOp(const std::string &type, const framework::VariableNameMap &inputs,
+       const framework::VariableNameMap &outputs,
+       const framework::AttributeMap &attrs)
+      : NetOp(type, inputs, outputs, attrs) {
+    PADDLE_ENFORCE(!Inputs("X").empty(),
+                   "Inputs(X) of FCOp should not be null.");
+    PADDLE_ENFORCE(!Inputs("W").empty(),
+                   "Inputs(W) of FCOp should not be null.");
+    PADDLE_ENFORCE(!Outputs("MulOut").empty(),
+                   "Outputs(MulOut) of FCOp should not be null.");
+    PADDLE_ENFORCE_NE(Output("Out"), framework::kEmptyVarName,
+                      "Output(Out) of FCOp should not be null.");
+
+    auto x = Inputs("X");
+    auto w = Inputs("W");
+    auto mul_out = Outputs("MulOut");
+    PADDLE_ENFORCE_EQ(
+        x.size(), w.size(),
+        "The size of inputs X(%d) should be the same as that of weights W(%d).",
+        x.size(), w.size());
+    PADDLE_ENFORCE_EQ(mul_out.size(), x.size(),
+                      "The size of intermediate mul_out(%d) should be the same "
+                      "as that of inputs X(%d).",
+                      mul_out.size(), x.size());
+
+    size_t n = x.size();
+    PADDLE_ENFORCE_GE(n, static_cast<size_t>(1),
+                      "The size of inputs X(%d) should be no less than 1.", n);
+
+    auto x_num_col_dims = Attr<std::vector<int>>("xNumColDims");
+
+    // Set all values or set no values (use the default value)
+    if (!x_num_col_dims.empty()) {
+      PADDLE_ENFORCE_EQ(x_num_col_dims.size(), n,
+                        "The size of attribute xNumColDims(%d) should be the "
+                        "same as that of inputs X(%d).",
+                        x_num_col_dims.size(), n);
+    } else {
+      x_num_col_dims.resize(n);
+      for (size_t i = 0; i < n; i++) {
+        x_num_col_dims[i] = 1;
+      }
+    }
+
+    // mul_out[i] = X[i] * W[i]
+    for (size_t i = 0; i < n; i++) {
+      framework::AttributeMap mul_attr;
+      mul_attr["x_num_col_dims"] = static_cast<int>(x_num_col_dims[i]);
+      mul_attr["y_num_col_dims"] = static_cast<int>(1);
+      AppendOp(
+          framework::OpRegistry::CreateOp("mul", {{"X", {x[i]}}, {"Y", {w[i]}}},
+                                          {{"Out", {mul_out[i]}}}, mul_attr));
+    }
+
+    // sum_out = X[0] * W[0] + ... + X[n-1] * W[n-1]
+    auto sum_out = mul_out[0];
+    if (n > 1) {
+      PADDLE_ENFORCE_NE(Output("SumOut"), framework::kEmptyVarName,
+                        "Output(SumOut) of FCOp should not be null when the "
+                        "size of Inputs(X) > 1.");
+
+      sum_out = Output("SumOut");
+      AppendOp(framework::OpRegistry::CreateOp("sum", {{"X", {mul_out}}},
+                                               {{"Out", {sum_out}}}, {}));
+    } else {
+      if (Output("SumOut") != framework::kEmptyVarName) {
+        this->Rename(Output("SumOut"), framework::kEmptyVarName);
+      }
+    }
+
+    // add_out = sum_out + b
+    auto b = Input("B");
+    auto add_out = sum_out;
+    if (b != framework::kEmptyVarName) {
+      PADDLE_ENFORCE_NE(
+          Output("AddOut"), framework::kEmptyVarName,
+          "Output(AddOut) of FCOp should not be null when Input(B) is set.");
+
+      add_out = Output("AddOut");
+      AppendOp(framework::OpRegistry::CreateOp(
+          "rowwise_add", {{"X", {sum_out}}, {"b", {Input("B")}}},
+          {{"Out", {add_out}}}, {}));
+    } else {
+      if (Output("AddOut") != framework::kEmptyVarName) {
+        this->Rename(Output("AddOut"), framework::kEmptyVarName);
+      }
+    }
+
+    auto activation = Attr<std::string>("activation");
+    AppendOp(framework::OpRegistry::CreateOp(activation, {{"X", {add_out}}},
+                                             {{"Y", {Output("Out")}}}, {}));
+    CompleteAddOp(false);
+  }
+};
+
+class FCOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  FCOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(A vector of Tensors) each input Tensor can be of arbitrary "
+             "dimension, and will be reshaped to a 2-D matrix of size "
+             "(minibatch, number_of_input_features) according to attribute "
+             "xNumColDims.")
+        .AsDuplicable();
+    AddInput("W",
+             "(A vector of Tensors) the weights of FC operator, a "
+             "vector of 2-D matrix of size "
+             "(number_of_input_features, number_of_neurons).")
+        .AsDuplicable();
+    AddInput("B",
+             "(Tensor) the bias of FC operator, a 1-D vector of size "
+             "number_of_neurons.");
+
+    AddOutput("Out",
+              "(Tensor) the activated output matrix of FC operator, a 2-D "
+              "matrix of size (minibatch, number_of_neurons).");
+    AddOutput("MulOut",
+              "(A vector of Tensors) the intermediate outputs of FC operator, "
+              "each Tensor saving the product of X_i * W_i.")
+        .AsIntermediate()
+        .AsDuplicable();
+    AddOutput(
+        "SumOut",
+        "(Tensor) the intermediate output of FC operator, "
+        "saving the sum of the products of X and W, that is sum{X_i * W_i}.")
+        .AsIntermediate();
+    AddOutput("AddOut",
+              "(Tensor) the non-actived output of FC operator, "
+              "saving sum{X_i * W_i} + B.")
+        .AsIntermediate();
+    AddAttr<std::string>(
+        "activation",
+        "(string, default identity) the activation type of FC operator.")
+        .SetDefault("identity")
+        .InEnum({"identity", "sigmoid", "softmax"});
+    AddAttr<std::vector<int>>(
+        "xNumColDims",
+        "(std::vector<int>) The inputs Tensors of FC operator can be of "
+        "more than 2 dimensions. In that case, each input Tensor `X_i` will be "
+        "reshaped to a 2-D matrix. The matrix's first dimension "
+        "(the length of column) will be the product of `X_i`'s last "
+        "`xNumColDims_i` dimensions, that is "
+        "`X_i.dims[0] x ... x X_i.dims[xNumColDims_i - 1]`. "
+        "The matrix's second dimension (the length of row) will be the product "
+        "of `X_i`'s first `rank - xNumColDims_i` dimensions, that is "
+        "`X_i.dims[xNumColDims_i] x ... x X_i.dims[rank - 1]`)")
+        .SetDefault(std::vector<int>{});
+
+    AddComment(R"DOC(
+Fully Connected Operator, known as Fully Connected Layer or Inner Product Layer
+in Convolutional Neural Networks. Neurons in a fully connected layer have
+full connections to all activations in the previous layer.
+It computes an inner product of a set of
+learned weights with a matrix multiplication followed by a bias offset
+(optionally).
+
+Equation:
+  Out = Act(sum_n{X_i * W_i} + B)
+
+where X_i is Tensor that will be reshaped to a 2-D matrix of size (M x K),
+usually M is the minibatch size and K is the number of input features.
+W_i is a 2-D matrix of size (K x N), where N means the number of neurons
+in the fully connected layer. B is a 1-D vector of size N.
+Thus, the output Out is a 2-D matrix of size (M x N).
+Activation type can be set to `identity` (default), `sigmoid` or `softmax`.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(fc, ops::FCOp, ops::FCOpMaker);

paddle/operators/fill_zeros_like_op.cc

@@ -23,6 +23,13 @@ class FillZerosLikeOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("Src"),
+        "Input(Src) of FillZerosLikeOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Dst"),
+        "Output(Dst) of FillZerosLikeOp should not be null.");
+
     ctx.Output<framework::LoDTensor>("Dst")->Resize(
         ctx.Input<framework::Tensor>("Src")->dims());
   }

paddle/operators/gather_op.cc

@@ -24,6 +24,13 @@ class GatherOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of GatherOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"),
+                            "Input(Index) of GatherOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of GatherOp should not be null.");
+
     int batch_size = ctx.Input<Tensor>("Index")->dims()[0];
     PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0");
     framework::DDim output_dims(ctx.Input<Tensor>("X")->dims());

paddle/operators/gaussian_random_op.cc

@@ -43,8 +43,12 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext& context) const override {
-    auto* tensor = context.Output<framework::LoDTensor>("Out");
+  void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of GaussianRandomOp should not be null.");
+
+    auto* tensor = ctx.Output<framework::LoDTensor>("Out");
     auto dims = Attr<std::vector<int>>("dims");
     std::vector<int64_t> temp;
     temp.reserve(dims.size());

paddle/operators/identity_op.cc

@@ -27,7 +27,7 @@ class IdentityOpMaker : public framework::OpProtoAndCheckerMaker {
                   framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input tensor of identity operator.");
-    AddOutput("Out", "The output tensor of identity operator.");
+    AddOutput("Y", "The output tensor of identity operator.");
     AddComment(R"DOC(
 The identity operator is an alias of the scale operator
 with the attribute scale fixed to 1.0.
@@ -42,9 +42,15 @@ class IdentityOp : public NetOp {
              const framework::VariableNameMap &outputs,
              const framework::AttributeMap &attrs)
       : NetOp(type, inputs, outputs, attrs) {
+    PADDLE_ENFORCE_NE(Input("X"), framework::kEmptyVarName,
+                      "Input(X) of IdentityOp should not be null.");
+    PADDLE_ENFORCE_NE(Output("Y"), framework::kEmptyVarName,
+                      "Output(Y) of IdentityOp should not be null.");
+
     AppendOp(framework::OpRegistry::CreateOp(
-        "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Out")}}},
+        "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Y")}}},
         {{"scale", static_cast<AttrType>(1)}}));
+    CompleteAddOp(false);
   }
 };
 

paddle/operators/lookup_table_op.cc

@@ -22,10 +22,17 @@ class LookupTableOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &context) const override {
-    auto table_t = context.Input<Tensor>("W");
-    auto ids_t = context.Input<Tensor>("Ids");
-    auto output_t = context.Output<framework::LoDTensor>("Out");
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("W"),
+                            "Input(W) of LookupTableOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Ids"),
+                            "Input(Ids) of LookupTableOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of LookupTableOp should not be null.");
+
+    auto table_t = ctx.Input<Tensor>("W");
+    auto ids_t = ctx.Input<Tensor>("Ids");
+    auto output_t = ctx.Output<framework::LoDTensor>("Out");
 
     output_t->Resize({ids_t->dims()[0], table_t->dims()[1]});
   }

paddle/operators/mean_op.cc

@@ -24,7 +24,9 @@ class MeanOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of MeanOp must be initialized.");
+                            "Input(X) of MeanOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of MeanOp should not be null.");
     ctx.Output<framework::LoDTensor>("Out")->Resize({1});
   }
 };

paddle/operators/minus_op.cc

@@ -27,6 +27,13 @@ class MinusOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of MinusOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of MinusOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of MinusOp should not be null.");
+
     auto *left_tensor = ctx.Input<framework::Tensor>("X");
     auto *right_tensor = ctx.Input<framework::Tensor>("Y");
 
@@ -64,7 +71,7 @@ class MinusGradOp : public NetOp {
 
     // x_grad = out_grad
     AppendOp(framework::OpRegistry::CreateOp("identity", {{"X", {out_grad}}},
-                                             {{"Out", {x_grad}}}, {}));
+                                             {{"Y", {x_grad}}}, {}));
 
     framework::AttributeMap scale_attr;
     scale_attr["scale"] = static_cast<AttrType>(-1);
@@ -77,8 +84,6 @@ class MinusGradOp : public NetOp {
 }  // namespace operators
 }  // namespace paddle
 
-USE_OP(scale);
-USE_NO_KERNEL_OP(identity);
 namespace ops = paddle::operators;
 REGISTER_OP(minus, ops::MinusOp, ops::MinusOpMaker, minus_grad,
             ops::MinusGradOp<float>);

paddle/operators/mul_op.cc

@@ -26,6 +26,13 @@ class MulOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of MulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of MulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of MulOp should not be null.");
+
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto y_dims = ctx.Input<Tensor>("Y")->dims();
     int x_num_col_dims = Attr<int>("x_num_col_dims");

paddle/operators/onehot_cross_entropy_op.cc

@@ -23,6 +23,16 @@ class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("X"),
+        "Input(X) of OnehotCrossEntropyOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("label"),
+        "Input(label) of OnehotCrossEntropyOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Y"),
+        "Output(Y) of OnehotCrossEntropyOp should not be null.");
+
     auto *X = ctx.Input<Tensor>("X");
     auto *label = ctx.Input<Tensor>("label");
 

paddle/operators/pad_op.cc

@@ -25,6 +25,11 @@ class PadOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of PadOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of PadOp should not be null.");
+
     auto x_dim = ctx.Input<Tensor>("X")->dims();
     auto paddings = Attr<std::vector<int>>("paddings");
     PADDLE_ENFORCE_EQ(x_dim.size() * 2, int64_t(paddings.size()),

paddle/operators/reshape_op.cc

@@ -28,7 +28,11 @@ class ReshapeOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     // input check
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) shouldn't be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of ReshapeOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ReshapeOp should not be null.");
+
     auto shape = ctx.Attr<std::vector<int>>("shape");
     PADDLE_ENFORCE(shape.size() > 0, "Attr(shape) shouldn't be empty.");
     for (auto dim : shape) {

paddle/operators/rowwise_add_op.cc

@@ -25,6 +25,13 @@ class RowwiseAddOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of RowwiseAddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("b"),
+                            "Input(b) of RowwiseAddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of RowwiseAddOp should not be null.");
+
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto b_dims = ctx.Input<Tensor>("b")->dims();
     PADDLE_ENFORCE_GT(

paddle/operators/scale_op.cc

@@ -27,6 +27,11 @@ class ScaleOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of ScaleOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ScaleOp should not be null.");
+
     auto *in = ctx.Input<framework::Tensor>("X");
     auto *out = ctx.Output<framework::LoDTensor>("Out");
     out->Resize(in->dims());

paddle/operators/scatter_op.cc

@@ -24,6 +24,15 @@ class ScatterOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Ref"),
+                            "Input(Ref) of ScatterOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"),
+                            "Input(Index) of ScatterOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Updates"),
+                            "Input(Updates) of ScatterOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ScatterOp should not be null.");
+
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("Index")->dims().size(), 1,
                       "Update Index should be 1-D.");
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("Ref")->dims().size(),

paddle/operators/sequence_avg_pool_op.cc

@@ -23,9 +23,12 @@ class SequenceAvgPoolOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of SequenceAvgPoolOp"
-                            "must be initialized.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("X"), "Input(X) of SequenceAvgPoolOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of SequenceAvgPoolOp should not be null.");
+
     auto* x = ctx.Input<framework::LoDTensor>("X");
     auto dims = x->dims();
     auto lod = x->lod();
@@ -60,7 +63,9 @@ class SequenceAvgPoolGradOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
-                            "Gradient of Out should not be null");
+                            "Gradient of Out should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "The input X should not be null.");
     auto og_dims =
         ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"))->dims();
     auto x_dims = ctx.Input<framework::LoDTensor>("X")->dims();

paddle/operators/sequence_avg_pool_op.h

@@ -21,6 +21,9 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 using LoDTensor = framework::LoDTensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
@@ -43,8 +46,8 @@ class SequenceAvgPoolKernel : public framework::OpKernel {
                                  static_cast<int>(lod[0][i + 1]));
       Tensor out_t = out->Slice<T>(i, i + 1);
       int64_t h = static_cast<int64_t>(lod[0][i + 1] - lod[0][i]);
-      auto in_e = EigenMatrix<T>::From(in_t, {h, w});
-      auto out_e = EigenMatrix<T>::From(out_t, {h, w});
+      auto in_e = EigenMatrix<T>::From(in_t, framework::make_ddim({h, w}));
+      auto out_e = EigenVector<T>::Flatten(out_t);
       out_e.device(place) = in_e.mean(Eigen::array<int, 1>({{0}}));
     }
   }
@@ -54,9 +57,9 @@ template <typename Place, typename T>
 class SequenceAvgPoolGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* in = context.Output<LoDTensor>("X");
-    auto* in_g = context.Output<LoDTensor>(framework::GradVarName("X"));
+    auto* in = context.Input<LoDTensor>("X");
     auto* out_g = context.Input<LoDTensor>(framework::GradVarName("Out"));
+    auto* in_g = context.Output<LoDTensor>(framework::GradVarName("X"));
 
     auto dims = in->dims();
     auto lod = in->lod();
@@ -71,7 +74,7 @@ class SequenceAvgPoolGradKernel : public framework::OpKernel {
       int64_t h = static_cast<int64_t>(lod[0][i + 1] - lod[0][i]);
       auto in_g_e = EigenMatrix<T>::From(in_g_t, {h, w});
       auto out_g_e = EigenMatrix<T>::From(out_g_t, {1, w});
-      Eigen::DSizes<int, 2> bcast(h, w);
+      Eigen::DSizes<int, 2> bcast(h, 1);
       in_g_e.device(place) = (out_g_e / static_cast<T>(h)).broadcast(bcast);
     }
   }

paddle/operators/sgd_op.cc

@@ -23,6 +23,13 @@ class SGDOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("param"),
+                            "Input(param) of SGDOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("grad"),
+                            "Input(grad) of SGDOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("param_out"),
+                            "Output(param_out) of SGDOp should not be null.");
+
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("param")->dims(),
                       ctx.Input<Tensor>("grad")->dims(),
                       "Two input of SGD Op's dimension must be same.");

paddle/operators/sigmoid_op.cc

@@ -23,6 +23,11 @@ class SigmoidOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of SigmoidOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
+                            "Output(Y) of SigmoidOp should not be null.");
+
     ctx.Output<framework::LoDTensor>("Y")->Resize(
         ctx.Input<Tensor>("X")->dims());
   }

paddle/operators/softmax_op.cc

@@ -23,6 +23,11 @@ class SoftmaxOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of SoftmaxOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
+                            "Output(Y) of SoftmaxOp should not be null.");
+
     PADDLE_ENFORCE(ctx.Input<Tensor>("X")->dims().size() == 2UL,
                    "The input of softmax op must be a matrix.");
     ctx.Output<framework::LoDTensor>("Y")->Resize(

paddle/operators/split_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+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/operators/split_op.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+using framework::Tensor;
+
+class SplitOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    // infershape
+    auto *in = ctx.Input<framework::Tensor>("X");
+    auto outs = ctx.MultiOutput<framework::LoDTensor>("Out");
+    size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
+    size_t num = static_cast<size_t>(ctx.Attr<int>("num"));
+    std::vector<int> sections =
+        static_cast<std::vector<int>>(ctx.Attr<std::vector<int>>("sections"));
+    const size_t n = outs.size();
+
+    if (num > 0) {
+      int64_t in_axis_dim = in->dims()[axis];
+      PADDLE_ENFORCE_EQ(in_axis_dim % num, 0,
+                        "tensor split does not result"
+                        " in an equal division");
+      size_t out_axis_dim = in_axis_dim / num;
+      for (size_t i = 0; i < n; ++i) {
+        auto dim = in->dims();
+        dim[axis] = out_axis_dim;
+        outs[i]->Resize(dim);
+      }
+    } else if (sections.size() > 0) {
+      PADDLE_ENFORCE_EQ(sections.size(), n,
+                        "tensor split sections size"
+                        "should be equal to output size.");
+      for (size_t i = 0; i < n; ++i) {
+        auto dim = in->dims();
+        dim[axis] = sections[i];
+        outs[i]->Resize(dim);
+      }
+    } else {
+      PADDLE_ENFORCE_NOT_NULL(nullptr, "split operator should",
+                              " specify indices or sections.");
+    }
+  }
+};
+
+class SplitOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SplitOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "the input tensor of split operator.");
+    AddOutput("Out", "the output tensors of split operator.").AsDuplicable();
+    AddComment(R"DOC(
+      Split the input tensor into multiple sub-tensors.
+      Example:
+        Input = [[1,2],
+                 [3,4],
+                 [5,6]]
+        sections = [2,1]
+        axis = 0
+        Output[0] = [[1,2],
+                     [3,4]]
+        Output[1] = [[5,6]]
+
+    )DOC");
+    AddAttr<std::vector<int>>("sections",
+                              "the length for each"
+                              "output along with the specify axis.")
+        .SetDefault(std::vector<int>{});
+    AddAttr<int>("num",
+                 "number of the sub-tensors, it must evenly divide "
+                 "Input.dims()[axis]")
+        .SetDefault(0);
+    AddAttr<int>("axis", "The axis which the input will be splited on.")
+        .SetDefault(0);
+  }
+};
+
+class SplitOpGrad : public NetOp {
+ public:
+  SplitOpGrad(const std::string &type, const framework::VariableNameMap &inputs,
+              const framework::VariableNameMap &outputs,
+              const framework::AttributeMap &attrs)
+      : NetOp(type, inputs, outputs, attrs) {
+    auto out_grad = Inputs(framework::GradVarName("Out"));
+    auto x_grad = Output(framework::GradVarName("X"));
+    AppendOp(framework::OpRegistry::CreateOp("concat", {{"X", out_grad}},
+                                             {{"Out", {x_grad}}}, attrs));
+    CompleteAddOp(false);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+USE_CPU_ONLY_OP(concat);
+REGISTER_OP(split, ops::SplitOp, ops::SplitOpMaker, split_grad,
+            ops::SplitOpGrad);
+REGISTER_OP_CPU_KERNEL(split,
+                       ops::SplitKernel<paddle::platform::CPUPlace, float>);

paddle/operators/split_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+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 <vector>
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class SplitKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* in = ctx.Input<framework::Tensor>("X");
+    auto outs = ctx.MultiOutput<framework::Tensor>("Out");
+    int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
+    size_t before = 1, after = 1;
+    const size_t n = outs.size();
+    size_t input_axis_dim = in->dims()[axis];
+
+    for (int64_t i = 0; i < in->dims().size(); ++i) {
+      if (i == axis) {
+        continue;
+      }
+      if (i < axis) {
+        before *= in->dims()[i];
+      } else {
+        after *= in->dims()[i];
+      }
+    }
+    size_t input_offset = 0;
+    for (size_t i = 0; i < n; i++) {
+      auto& out = outs[i];
+      size_t axis_dim = out->dims()[axis];
+      for (size_t j = 0; j < before; j++) {
+        size_t len = axis_dim * after * sizeof(T);
+        T* dest =
+            out->mutable_data<T>(platform::CPUPlace()) + axis_dim * after * j;
+        const T* src =
+            in->data<T>() + input_offset + input_axis_dim * after * j;
+        memcpy(dest, src, len);
+      }
+      input_offset += axis_dim * after;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/squared_l2_distance_op.cc

@@ -23,12 +23,18 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of SquaredL2DistanceOp "
-                            "must be initialized.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
-                            "Target of SquaredL2DistanceOp "
-                            "must be initialized.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("X"),
+        "Input(X) of SquaredL2DistanceOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("Y"),
+        "Input(Y) of SquaredL2DistanceOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("sub_result"),
+        "Output(sub_result) of SquaredL2DistanceOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of SquaredL2DistanceOp should not be null.");
 
     auto* x = ctx.Input<Tensor>("X");
     auto x_dims = x->dims();

paddle/operators/sum_op.cc

@@ -22,6 +22,11 @@ class SumOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE(!ctx.MultiInputVar("X").empty(),
+                   "Input(X) of SumOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of SumOp should not be null.");
+
     auto ins = ctx.MultiInput<framework::Tensor>("X");
     auto *out = ctx.Output<framework::LoDTensor>("Out");
     int N = ins.size();

paddle/operators/top_k_op.cc

@@ -24,7 +24,12 @@ class TopkOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of TopkOP must be initialized.");
+                            "Input(X) of TopkOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of TopkOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Indices"),
+                            "Output(Indices) of TopkOp should not be null.");
+
     auto *input = ctx.Input<framework::Tensor>("X");
     const int k = static_cast<int>(ctx.Attr<int>("k"));
 

paddle/operators/uniform_random_op.cc

@@ -48,6 +48,10 @@ class UniformRandomOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of UniformRandomOp should not be null.");
+
     PADDLE_ENFORCE(Attr<float>("min") < Attr<float>("max"),
                    "uniform_random's min must less then max");
     auto* tensor = ctx.Output<framework::LoDTensor>("Out");

paddle/platform/CMakeLists.txt

@@ -24,3 +24,4 @@ cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator
 nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info)
 
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
+nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place)

paddle/platform/cuda_helper.h

@@ -24,6 +24,11 @@ namespace platform {
 #define USE_CUDA_ATOMIC(op, T) \
   CUDA_ATOMIC_WRAPPER(op, T) { return atomic##op(address, val); }
 
+// Default thread count per block(or block size).
+// TODO(typhoonzero): need to benchmark against setting this value
+//                    to 1024.
+constexpr int PADDLE_CUDA_NUM_THREADS = 512;
+
 // For atomicAdd.
 USE_CUDA_ATOMIC(Add, float);
 

paddle/platform/details/device_ptr_cast.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#ifndef __NVCC__
+#error device_ptr_cast must be include by .cu file
+#endif
+
+#include <thrust/device_ptr.h>
+
+namespace paddle {
+namespace platform {
+namespace details {
+template <typename T, bool is_ptr>
+struct DevicePtrCast;
+
+template <typename T>
+struct DevicePtrCast<T, true> {
+  using ELEM = typename std::remove_pointer<T>::type;
+  using RTYPE = thrust::device_ptr<ELEM>;
+
+  inline thrust::device_ptr<ELEM> operator()(ELEM* ele) const {
+    return thrust::device_pointer_cast(ele);
+  }
+};
+
+template <typename T>
+struct DevicePtrCast<T, false> {
+  using RTYPE = T;
+  inline RTYPE operator()(RTYPE it) const { return it; }
+};
+
+// Cast T to thrust::device_ptr if T is a pointer.
+// Otherwise, e.g., T is a iterator, return T itself.
+template <typename T>
+auto DevPtrCast(T t) ->
+    typename DevicePtrCast<T, std::is_pointer<T>::value>::RTYPE {
+  DevicePtrCast<T, std::is_pointer<T>::value> cast;
+  return cast(t);
+}
+
+}  // namespace details
+}  // namespace platform
+}  // namespace paddle

paddle/platform/transform.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+
+#include "paddle/platform/enforce.h"
+#include "paddle/platform/hostdevice.h"
+#include "paddle/platform/place.h"
+
+#include <algorithm>
+#include <type_traits>
+#ifdef __NVCC__
+#include <thrust/transform.h>
+#include "paddle/platform/details/device_ptr_cast.h"
+#endif
+
+namespace paddle {
+namespace platform {
+// Transform on host or device. It provides the same API in std library.
+template <typename Place, typename InputIter, typename OutputIter,
+          typename UnaryOperation>
+void Transform(Place place, InputIter first, InputIter last, OutputIter result,
+               UnaryOperation op) {
+  if (is_cpu_place(place)) {
+    std::transform(first, last, result, op);
+  } else {
+#ifdef __NVCC__
+    using namespace details;
+    thrust::transform(DevPtrCast(first), DevPtrCast(last), DevPtrCast(result),
+                      op);
+#else
+    PADDLE_THROW("Do not invoke `Transform<GPUPlace>` in .cc file");
+#endif
+  }
+}
+
+template <typename Place, typename InputIter1, typename InputIter2,
+          typename OutputIter, typename BinaryOperation>
+void Transform(Place place, InputIter1 first1, InputIter1 last1,
+               InputIter2 first2, OutputIter result, BinaryOperation op) {
+  if (is_cpu_place(place)) {
+    std::transform(first1, last1, first2, result, op);
+  } else {
+#ifdef __NVCC__
+    using namespace details;
+    thrust::transform(DevPtrCast(first1), DevPtrCast(last1), DevPtrCast(first2),
+                      DevPtrCast(result), op);
+#else
+    PADDLE_THROW("Do not invoke `Transform<GPUPlace>` in .cc file");
+#endif
+  }
+};
+
+}  // namespace platform
+}  // namespace paddle

paddle/platform/transform_test.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include <gtest/gtest.h>
+#include "paddle/memory/memcpy.h"
+#include "paddle/memory/memory.h"
+#include "paddle/platform/transform.h"
+
+template <typename T>
+class Scale {
+ public:
+  explicit Scale(const T& scale) : scale_(scale) {}
+
+  HOSTDEVICE T operator()(const T& a) const { return a * scale_; }
+
+ private:
+  T scale_;
+};
+
+template <typename T>
+class Multiply {
+ public:
+  HOSTDEVICE T operator()(const T& a, const T& b) const { return a * b; }
+};
+
+TEST(Transform, CPUUnary) {
+  using namespace paddle::platform;
+  float buf[4] = {0.1, 0.2, 0.3, 0.4};
+  Transform(CPUPlace(), buf, buf + 4, buf, Scale<float>(10));
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_NEAR(buf[i], static_cast<float>(i + 1), 1e-5);
+  }
+}
+
+TEST(Transform, GPUUnary) {
+  using namespace paddle::platform;
+  using namespace paddle::memory;
+  GPUPlace gpu0(0);
+  float cpu_buf[4] = {0.1, 0.2, 0.3, 0.4};
+  float* gpu_buf = static_cast<float*>(Alloc(gpu0, sizeof(float) * 4));
+  Copy(gpu0, gpu_buf, CPUPlace(), cpu_buf, sizeof(cpu_buf));
+  Transform(gpu0, gpu_buf, gpu_buf + 4, gpu_buf, Scale<float>(10));
+  Copy(CPUPlace(), cpu_buf, gpu0, gpu_buf, sizeof(cpu_buf));
+  Free(gpu0, gpu_buf);
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_NEAR(cpu_buf[i], static_cast<float>(i + 1), 1e-5);
+  }
+}
+
+TEST(Transform, CPUBinary) {
+  using namespace paddle::platform;
+  using namespace paddle::memory;
+  int buf[4] = {1, 2, 3, 4};
+  Transform(CPUPlace(), buf, buf + 4, buf, buf, Multiply<int>());
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_EQ((i + 1) * (i + 1), buf[i]);
+  }
+}
+
+TEST(Transform, GPUBinary) {
+  using namespace paddle::platform;
+  using namespace paddle::memory;
+  int buf[4] = {1, 2, 3, 4};
+  GPUPlace gpu0(0);
+  int* gpu_buf = static_cast<int*>(Alloc(gpu0, sizeof(buf)));
+  Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf));
+  Transform(gpu0, gpu_buf, gpu_buf + 4, gpu_buf, gpu_buf, Multiply<int>());
+  Copy(CPUPlace(), buf, gpu0, gpu_buf, sizeof(buf));
+  Free(gpu0, gpu_buf);
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_EQ((i + 1) * (i + 1), buf[i]);
+  }
+}
\ No newline at end of file

paddle/pybind/CMakeLists.txt

 if(WITH_PYTHON)
-cc_library(paddle_pybind SHARED
+  cc_library(paddle_pybind SHARED
     SRCS pybind.cc
     DEPS pybind python backward
     ${GLOB_OP_LIB})

paddle/pybind/pybind.cc

@@ -19,6 +19,7 @@ limitations under the License. */
 #include "paddle/framework/backward.h"
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/cond_op.h"
 #include "paddle/operators/net_op.h"
 #include "paddle/operators/recurrent_op.h"
 #include "paddle/platform/enforce.h"
@@ -288,6 +289,28 @@ All parameter, weight, gradient are variables in Paddle.
            [](operators::RecurrentOp &self, const operators::NetOp &net)
                -> void { self.set_stepnet(net.Clone()); });
 
+  // cond_op
+  py::class_<operators::CondOp, OperatorBase>(m, "CondOp")
+      .def_static("create",
+                  [](py::bytes protobin) -> operators::CondOp * {
+                    OpDesc desc;
+                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                                   "Cannot parse user input to OpDesc");
+                    PADDLE_ENFORCE(desc.IsInitialized(),
+                                   "User OpDesc is not initialized, reason %s",
+                                   desc.InitializationErrorString());
+                    auto cond_op = OpRegistry::CreateOp(desc);
+                    return static_cast<operators::CondOp *>(cond_op.release());
+                  })
+      .def("set_truenet",
+           [](operators::CondOp &self, const operators::NetOp &net) -> void {
+             self.set_truenet(net.Clone());
+           })
+      .def("set_falsenet",
+           [](operators::CondOp &self, const operators::NetOp &net) -> void {
+             self.set_falsenet(net.Clone());
+           });
+
   m.def("unique_integer", UniqueIntegerGenerator);
 
   m.def("is_compile_gpu", IsCompileGPU);

python/paddle/trainer_config_helpers/networks.py

@@ -11,10 +11,8 @@
 # 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.
-"""
 
-"""
-# from activations import *
+
 from activations import LinearActivation, ReluActivation, SoftmaxActivation, \
     IdentityActivation, TanhActivation, SequenceSoftmaxActivation
 from attrs import ExtraAttr
@@ -55,49 +53,49 @@ def sequence_conv_pool(input,
                        context_attr=None,
                        pool_attr=None):
     """
-    Text convolution pooling layers helper.
+    Text convolution pooling group.
 
     Text input => Context Projection => FC Layer => Pooling => Output.
 
-    :param name: name of output layer(pooling layer name)
+    :param name: group name.
     :type name: basestring
-    :param input: name of input layer
+    :param input: input layer.
     :type input: LayerOutput
     :param context_len: context projection length. See
                         context_projection's document.
     :type context_len: int
     :param hidden_size: FC Layer size.
     :type hidden_size: int
-    :param context_start: context projection length. See
+    :param context_start: context start position. See
                           context_projection's context_start.
-    :type context_start: int or None
+    :type context_start: int|None
     :param pool_type: pooling layer type. See pooling_layer's document.
-    :type pool_type: BasePoolingType.
+    :type pool_type: BasePoolingType
     :param context_proj_layer_name: context projection layer name.
                                     None if user don't care.
     :type context_proj_layer_name: basestring
-    :param context_proj_param_attr: context projection parameter attribute.
-                                    None if user don't care.
-    :type context_proj_param_attr: ParameterAttribute or None.
+    :param context_proj_param_attr: padding parameter attribute of context projection layer.
+                                    If false, it means padding always be zero.
+    :type context_proj_param_attr: ParameterAttribute|None
     :param fc_layer_name: fc layer name. None if user don't care.
     :type fc_layer_name: basestring
     :param fc_param_attr: fc layer parameter attribute. None if user don't care.
-    :type fc_param_attr: ParameterAttribute or None
+    :type fc_param_attr: ParameterAttribute|None
     :param fc_bias_attr: fc bias parameter attribute. False if no bias,
                          None if user don't care.
-    :type fc_bias_attr: ParameterAttribute or None
-    :param fc_act: fc layer activation type. None means tanh
+    :type fc_bias_attr: ParameterAttribute|False|None
+    :param fc_act: fc layer activation type. None means tanh.
     :type fc_act: BaseActivation
-    :param pool_bias_attr: pooling layer bias attr. None if don't care.
-                           False if no bias.
-    :type pool_bias_attr: ParameterAttribute or None.
+    :param pool_bias_attr: pooling layer bias attr. False if no bias.
+                           None if user don't care.
+    :type pool_bias_attr: ParameterAttribute|False|None
     :param fc_attr: fc layer extra attribute.
     :type fc_attr: ExtraLayerAttribute
     :param context_attr: context projection layer extra attribute.
     :type context_attr: ExtraLayerAttribute
     :param pool_attr: pooling layer extra attribute.
     :type pool_attr: ExtraLayerAttribute
-    :return: output layer name.
+    :return: layer's output.
     :rtype: LayerOutput
     """
     # Set Default Value to param
@@ -163,45 +161,45 @@ def simple_img_conv_pool(input,
     """
     Simple image convolution and pooling group.
 
-    Input => conv => pooling
+    Img input => Conv => Pooling => Output.
 
-    :param name: group name
+    :param name: group name.
     :type name: basestring
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
-    :param filter_size: see img_conv_layer for details
+    :param filter_size: see img_conv_layer for details.
     :type filter_size: int
-    :param num_filters: see img_conv_layer for details
+    :param num_filters: see img_conv_layer for details.
     :type num_filters: int
-    :param pool_size: see img_pool_layer for details
+    :param pool_size: see img_pool_layer for details.
     :type pool_size: int
-    :param pool_type: see img_pool_layer for details
+    :param pool_type: see img_pool_layer for details.
     :type pool_type: BasePoolingType
-    :param act: see img_conv_layer for details
+    :param act: see img_conv_layer for details.
     :type act: BaseActivation
-    :param groups: see img_conv_layer for details
+    :param groups: see img_conv_layer for details.
     :type groups: int
-    :param conv_stride: see img_conv_layer for details
+    :param conv_stride: see img_conv_layer for details.
     :type conv_stride: int
-    :param conv_padding: see img_conv_layer for details
+    :param conv_padding: see img_conv_layer for details.
     :type conv_padding: int
-    :param bias_attr: see img_conv_layer for details
+    :param bias_attr: see img_conv_layer for details.
     :type bias_attr: ParameterAttribute
-    :param num_channel: see img_conv_layer for details
+    :param num_channel: see img_conv_layer for details.
     :type num_channel: int
-    :param param_attr: see img_conv_layer for details
+    :param param_attr: see img_conv_layer for details.
     :type param_attr: ParameterAttribute
-    :param shared_bias: see img_conv_layer for details
+    :param shared_bias: see img_conv_layer for details.
     :type shared_bias: bool
-    :param conv_layer_attr: see img_conv_layer for details
+    :param conv_layer_attr: see img_conv_layer for details.
     :type conv_layer_attr: ExtraLayerAttribute
-    :param pool_stride: see img_pool_layer for details
+    :param pool_stride: see img_pool_layer for details.
     :type pool_stride: int
-    :param pool_padding: see img_pool_layer for details
+    :param pool_padding: see img_pool_layer for details.
     :type pool_padding: int
-    :param pool_layer_attr: see img_pool_layer for details
+    :param pool_layer_attr: see img_pool_layer for details.
     :type pool_layer_attr: ExtraLayerAttribute
-    :return: Layer's output
+    :return: layer's output
     :rtype: LayerOutput
     """
     _conv_ = img_conv_layer(
@@ -253,47 +251,51 @@ def img_conv_bn_pool(input,
     """
     Convolution, batch normalization, pooling group.
     
-    :param name: group name
+    Img input => Conv => BN => Pooling => Output.
+
+    :param name: group name.
     :type name: basestring
-    :param input: layer's input
+    :param input: input layer.
     :type input: LayerOutput 
-    :param filter_size: see img_conv_layer's document
+    :param filter_size: see img_conv_layer for details.
     :type filter_size: int
-    :param num_filters: see img_conv_layer's document
+    :param num_filters: see img_conv_layer for details.
     :type num_filters: int
-    :param pool_size: see img_pool_layer's document.
+    :param pool_size: see img_pool_layer for details.
     :type pool_size: int
-    :param pool_type: see img_pool_layer's document.
+    :param pool_type: see img_pool_layer for details.
     :type pool_type: BasePoolingType
-    :param act: see batch_norm_layer's document.
+    :param act: see batch_norm_layer for details.
     :type act: BaseActivation
-    :param groups: see img_conv_layer's document
+    :param groups: see img_conv_layer for details.
     :type groups: int
-    :param conv_stride: see img_conv_layer's document.
+    :param conv_stride: see img_conv_layer for details.
     :type conv_stride: int
-    :param conv_padding: see img_conv_layer's document.
+    :param conv_padding: see img_conv_layer for details.
     :type conv_padding: int
-    :param conv_bias_attr: see img_conv_layer's document.
+    :param conv_bias_attr: see img_conv_layer for details.
     :type conv_bias_attr: ParameterAttribute
-    :param num_channel: see img_conv_layer's document.
+    :param num_channel: see img_conv_layer for details.
     :type num_channel: int
-    :param conv_param_attr: see img_conv_layer's document.
+    :param conv_param_attr: see img_conv_layer for details.
     :type conv_param_attr: ParameterAttribute
-    :param shared_bias: see img_conv_layer's document.
+    :param shared_bias: see img_conv_layer for details.
     :type shared_bias: bool
-    :param conv_layer_attr: see img_conv_layer's document.
+    :param conv_layer_attr: see img_conv_layer for details.
     :type conv_layer_attr: ExtraLayerOutput
-    :param bn_param_attr: see batch_norm_layer's document.
-    :type bn_param_attr: ParameterAttribute.
-    :param bn_bias_attr: see batch_norm_layer's document.
-    :param bn_layer_attr: ParameterAttribute.
-    :param pool_stride: see img_pool_layer's document.
+    :param bn_param_attr: see batch_norm_layer for details.
+    :type bn_param_attr: ParameterAttribute
+    :param bn_bias_attr: see batch_norm_layer for details.
+    :type bn_bias_attr: ParameterAttribute
+    :param bn_layer_attr: see batch_norm_layer for details.
+    :type bn_layer_attr: ExtraLayerAttribute
+    :param pool_stride: see img_pool_layer for details.
     :type pool_stride: int
-    :param pool_padding: see img_pool_layer's document.
+    :param pool_padding: see img_pool_layer for details.
     :type pool_padding: int
-    :param pool_layer_attr: see img_pool_layer's document.
+    :param pool_layer_attr: see img_pool_layer for details.
     :type pool_layer_attr: ExtraLayerAttribute
-    :return: Layer groups output
+    :return: layer's output
     :rtype: LayerOutput
     """
     __conv__ = img_conv_layer(
@@ -348,10 +350,10 @@ def img_conv_group(input,
     :param conv_batchnorm_drop_rate: if conv_with_batchnorm[i] is true,
         conv_batchnorm_drop_rate[i] represents the drop rate of each batch norm.
     :type conv_batchnorm_drop_rate: list
-    :param input: layer's input.
+    :param input: input layer.
     :type input: LayerOutput
-    :param conv_num_filter: output channels num.
-    :type conv_num_filter: int
+    :param conv_num_filter: list of output channels num.
+    :type conv_num_filter: list|tuple
     :param pool_size: pooling filter size.
     :type pool_size: int
     :param num_channels: input channels num.
@@ -362,18 +364,18 @@ def img_conv_group(input,
     :type conv_filter_size: int
     :param conv_act: activation funciton after convolution.
     :type conv_act: BaseActivation
-    :param conv_with_batchnorm: conv_with_batchnorm[i] represents
-        if there is a batch normalization after each convolution.
+    :param conv_with_batchnorm: if conv_with_batchnorm[i] is true,
+        there is a batch normalization operation after each convolution.
     :type conv_with_batchnorm: list
     :param pool_stride: pooling stride size.
     :type pool_stride: int
     :param pool_type: pooling type.
     :type pool_type: BasePoolingType
-    :param param_attr: Convolution param attribute.
+    :param param_attr: param attribute of convolution layer,
                        None means default attribute.
     :type param_attr: ParameterAttribute
-    :return: Layer's output
-    :type: LayerOutput
+    :return: layer's output
+    :rtype: LayerOutput
     """
     tmp = input
 
@@ -466,12 +468,14 @@ def vgg_16_network(input_image, num_channels, num_classes=1000):
     """
     Same model from https://gist.github.com/ksimonyan/211839e770f7b538e2d8
 
-    :param num_classes:
-    :param input_image:
+    :param num_classes: number of class.
+    :type num_classes: int
+    :param input_image: input layer.
     :type input_image: LayerOutput
-    :param num_channels:
+    :param num_channels: input channels num.
     :type num_channels: int
-    :return:
+    :return: layer's output
+    :rtype: LayerOutput
     """
 
     tmp = img_conv_group(
@@ -560,8 +564,8 @@ def simple_lstm(input,
     """
     Simple LSTM Cell.
 
-    It just combine a mixed layer with fully_matrix_projection and a lstmemory
-    layer. The simple lstm cell was implemented as follow equations.
+    It just combines a mixed layer with fully_matrix_projection and a lstmemory
+    layer. The simple lstm cell was implemented with follow equations.
 
     ..  math::
 
@@ -575,37 +579,37 @@ def simple_lstm(input,
 
         h_t & = o_t tanh(c_t)
 
-    Please refer **Generating Sequences With Recurrent Neural Networks** if you
-    want to know what lstm is. Link_ is here.
+    Please refer to **Generating Sequences With Recurrent Neural Networks** for more
+    details about lstm. Link_ is here.
 
     .. _Link: http://arxiv.org/abs/1308.0850
 
     :param name: lstm layer name.
     :type name: basestring
-    :param input: input layer name.
+    :param input: layer's input.
     :type input: LayerOutput
     :param size: lstm layer size.
     :type size: int
-    :param reverse: whether to process the input data in a reverse order
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param mat_param_attr: mixed layer's matrix projection parameter attribute.
+    :param mat_param_attr: parameter attribute of matrix projection in mixed layer.
     :type mat_param_attr: ParameterAttribute
     :param bias_param_attr: bias parameter attribute. False means no bias, None
                             means default bias.
     :type bias_param_attr: ParameterAttribute|False
-    :param inner_param_attr: lstm cell parameter attribute.
+    :param inner_param_attr: parameter attribute of lstm cell.
     :type inner_param_attr: ParameterAttribute
-    :param act: lstm final activiation type
+    :param act: last activiation type of lstm.
     :type act: BaseActivation
-    :param gate_act: lstm gate activiation type
+    :param gate_act: gate activiation type of lstm.
     :type gate_act: BaseActivation
-    :param state_act: lstm state activiation type.
+    :param state_act: state activiation type of lstm.
     :type state_act: BaseActivation
-    :param mixed_layer_attr: mixed layer's extra attribute.
+    :param mixed_layer_attr: extra attribute of mixed layer.
     :type mixed_layer_attr: ExtraLayerAttribute
-    :param lstm_cell_attr: lstm layer's extra attribute.
+    :param lstm_cell_attr: extra attribute of lstm.
     :type lstm_cell_attr: ExtraLayerAttribute
-    :return: lstm layer name.
+    :return: layer's output.
     :rtype: LayerOutput
     """
     fc_name = 'lstm_transform_%s' % name
@@ -643,9 +647,9 @@ def lstmemory_unit(input,
                    lstm_bias_attr=None,
                    lstm_layer_attr=None):
     """
-    Define calculations that a LSTM unit performs during a single time step.
-    This function itself is not a recurrent layer, so it can not be
-    directly used to process sequence inputs. This function is always used in
+    lstmemory_unit defines the caculation process of a LSTM unit during a 
+    single time step. This function is not a recurrent layer, so it can not be
+    directly used to process sequence input. This function is always used in
     recurrent_group (see layers.py for more details) to implement attention
     mechanism.
 
@@ -676,7 +680,7 @@ def lstmemory_unit(input,
                                    state_act=TanhActivation())
 
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param out_memory: output of previous time step
     :type out_memory: LayerOutput | None
@@ -684,15 +688,15 @@ def lstmemory_unit(input,
     :type name: basestring
     :param size: lstmemory unit size.
     :type size: int
-    :param param_attr: Parameter config, None if use default.
+    :param param_attr: parameter attribute, None means default attribute.
     :type param_attr: ParameterAttribute
-    :param act: lstm final activiation type
+    :param act: last activiation type of lstm.
     :type act: BaseActivation
-    :param gate_act: lstm gate activiation type
+    :param gate_act: gate activiation type of lstm.
     :type gate_act: BaseActivation
-    :param state_act: lstm state activiation type.
+    :param state_act: state activiation type of lstm.
     :type state_act: BaseActivation
-    :param input_proj_bias_attr: bias attribute for input-to-hidden projection.
+    :param input_proj_bias_attr: bias attribute for input to hidden projection.
                 False means no bias, None means default bias.
     :type input_proj_bias_attr: ParameterAttribute|False|None
     :param input_proj_layer_attr: extra layer attribute for input to hidden
@@ -700,8 +704,8 @@ def lstmemory_unit(input,
     :type input_proj_layer_attr: ExtraLayerAttribute
     :param lstm_bias_attr: bias parameter attribute of lstm layer.
                 False means no bias, None means default bias.
-    :type lstm_bias_attr: ParameterAttribute|False
-    :param lstm_layer_attr: lstm layer's extra attribute.
+    :type lstm_bias_attr: ParameterAttribute|False|None
+    :param lstm_layer_attr: extra attribute of lstm layer.
     :type lstm_layer_attr: ExtraLayerAttribute
     :return: lstmemory unit name.
     :rtype: LayerOutput
@@ -758,9 +762,9 @@ def lstmemory_group(input,
     lstm_group is a recurrent_group version of Long Short Term Memory. It
     does exactly the same calculation as the lstmemory layer (see lstmemory in
     layers.py for the maths) does. A promising benefit is that LSTM memory
-    cell states, or hidden states in every time step are accessible to the
+    cell states(or hidden states) in every time step are accessible to the
     user. This is especially useful in attention model. If you do not need to
-    access the internal states of the lstm, but merely use its outputs,
+    access the internal states of the lstm and merely use its outputs,
     it is recommended to use the lstmemory, which is relatively faster than
     lstmemory_group.
 
@@ -781,28 +785,28 @@ def lstmemory_group(input,
                                     gate_act=SigmoidActivation(),
                                     state_act=TanhActivation())
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param size: lstmemory group size.
     :type size: int
-    :param name: name of the lstmemory group.
+    :param name: name of lstmemory group.
     :type name: basestring
-    :param out_memory: output of previous time step
+    :param out_memory: output of previous time step.
     :type out_memory: LayerOutput | None
-    :param reverse: is lstm reversed
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param param_attr: Parameter config, None if use default.
+    :param param_attr: parameter attribute, None means default attribute.
     :type param_attr: ParameterAttribute
-    :param act: lstm final activiation type
+    :param act: last activiation type of lstm.
     :type act: BaseActivation
-    :param gate_act: lstm gate activiation type
+    :param gate_act: gate activiation type of lstm.
     :type gate_act: BaseActivation
-    :param state_act: lstm state activiation type.
+    :param state_act: state activiation type of lstm.
     :type state_act: BaseActivation
     :param lstm_bias_attr: bias parameter attribute of lstm layer.
                            False means no bias, None means default bias.
-    :type lstm_bias_attr: ParameterAttribute|False
-    :param input_proj_bias_attr: bias attribute for input-to-hidden projection.
+    :type lstm_bias_attr: ParameterAttribute|False|None
+    :param input_proj_bias_attr: bias attribute for input to hidden projection.
                 False means no bias, None means default bias.
     :type input_proj_bias_attr: ParameterAttribute|False|None
     :param input_proj_layer_attr: extra layer attribute for input to hidden
@@ -848,15 +852,15 @@ def gru_unit(input,
              gru_layer_attr=None,
              naive=False):
     """
-    Define calculations that a gated recurrent unit performs in a single time
-    step. This function itself is not a recurrent layer, so it can not be
-    directly used to process sequence inputs. This function is always used in
+    gru_unit defines the calculation process of a gated recurrent unit during a single 
+    time step. This function is not a recurrent layer, so it can not be
+    directly used to process sequence input. This function is always used in
     the recurrent_group (see layers.py for more details) to implement attention
     mechanism.
 
     Please see grumemory in layers.py for the details about the maths.
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param memory_boot: the initialization state of the LSTM cell.
     :type memory_boot: LayerOutput | None
@@ -864,12 +868,12 @@ def gru_unit(input,
     :type name: basestring
     :param size: hidden size of the gru.
     :type size: int
-    :param act: type of the activation
+    :param act: activation type of gru
     :type act: BaseActivation
-    :param gate_act: type of the gate activation
+    :param gate_act: gate activation type or gru
     :type gate_act: BaseActivation
-    :param gru_layer_attr: Extra parameter attribute of the gru layer.
-    :type gru_layer_attr: ParameterAttribute|False
+    :param gru_layer_attr: Extra attribute of the gru layer.
+    :type gru_layer_attr: ExtraLayerAttribute
     :return: the gru output layer.
     :rtype: LayerOutput
     """
@@ -915,7 +919,7 @@ def gru_group(input,
     does exactly the same calculation as the grumemory layer does. A promising
     benefit is that gru hidden states are accessible to the user. This is
     especially useful in attention model. If you do not need to access
-    any internal state, but merely use the outputs of a GRU, it is recommended
+    any internal state and merely use the outputs of a GRU, it is recommended
     to use the grumemory, which is relatively faster.
 
     Please see grumemory in layers.py for more detail about the maths.
@@ -924,12 +928,12 @@ def gru_group(input,
 
     ..  code-block:: python
 
-        gru = gur_group(input=[layer1],
+        gru = gru_group(input=[layer1],
                         size=256,
                         act=TanhActivation(),
                         gate_act=SigmoidActivation())
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param memory_boot: the initialization state of the LSTM cell.
     :type memory_boot: LayerOutput | None
@@ -937,16 +941,17 @@ def gru_group(input,
     :type name: basestring
     :param size: hidden size of the gru.
     :type size: int
-    :param reverse: whether to process the input data in a reverse order
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param act: type of the activiation
+    :param act: activiation type of gru
     :type act: BaseActivation
-    :param gate_act: type of the gate activiation
+    :param gate_act: gate activiation type of gru
     :type gate_act: BaseActivation
-    :param gru_bias_attr: bias. False means no bias, None means default bias.
-    :type gru_bias_attr: ParameterAttribute|False
-    :param gru_layer_attr: Extra parameter attribute of the gru layer.
-    :type gru_layer_attr: ParameterAttribute|False
+    :param gru_bias_attr: bias parameter attribute of gru layer,
+                          False means no bias, None means default bias.
+    :type gru_bias_attr: ParameterAttribute|False|None
+    :param gru_layer_attr: Extra attribute of the gru layer.
+    :type gru_layer_attr: ExtraLayerAttribute
     :return: the gru group.
     :rtype: LayerOutput
     """
@@ -986,11 +991,11 @@ def simple_gru(input,
                gru_layer_attr=None,
                naive=False):
     """
-    You maybe see gru_step_layer, grumemory in layers.py, gru_unit, gru_group,
+    You may see gru_step_layer, grumemory in layers.py, gru_unit, gru_group,
     simple_gru in network.py. The reason why there are so many interfaces is
     that we have two ways to implement recurrent neural network. One way is to
     use one complete layer to implement rnn (including simple rnn, gru and lstm)
-    with multiple time steps, such as recurrent_layer, lstmemory, grumemory. But,
+    with multiple time steps, such as recurrent_layer, lstmemory, grumemory. But 
     the multiplication operation :math:`W x_t` is not computed in these layers.
     See details in their interfaces in layers.py.
     The other implementation is to use an recurrent group which can ensemble a
@@ -1018,22 +1023,23 @@ def simple_gru(input,
 
         gru = simple_gru(input=[layer1], size=256)
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param name: name of the gru group.
     :type name: basestring
     :param size: hidden size of the gru.
     :type size: int
-    :param reverse: whether to process the input data in a reverse order
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param act: type of the activiation
+    :param act: activiation type of gru
     :type act: BaseActivation
-    :param gate_act: type of the gate activiation
+    :param gate_act: gate activiation type of gru
     :type gate_act: BaseActivation
-    :param gru_bias_attr: bias. False means no bias, None means default bias.
-    :type gru_bias_attr: ParameterAttribute|False
-    :param gru_layer_attr: Extra parameter attribute of the gru layer.
-    :type gru_layer_attr: ParameterAttribute|False
+    :param gru_bias_attr: bias parameter attribute of gru layer,
+                          False means no bias, None means default bias.
+    :type gru_bias_attr: ParameterAttribute|False|None
+    :param gru_layer_attr: Extra attribute of the gru layer.
+    :type gru_layer_attr: ExtraLayerAttribute
     :return: the gru group.
     :rtype: LayerOutput
     """
@@ -1071,8 +1077,8 @@ def simple_gru2(input,
                 mixed_layer_attr=None,
                 gru_cell_attr=None):
     """
-    simple_gru2 is the same with simple_gru, but using grumemory instead
-    Please see grumemory in layers.py for more detail about the maths.
+    simple_gru2 is the same with simple_gru, but using grumemory instead.
+    Please refer to grumemory in layers.py for more detail about the math.
     simple_gru2 is faster than simple_gru.
 
     The example usage is:
@@ -1081,22 +1087,23 @@ def simple_gru2(input,
 
         gru = simple_gru2(input=[layer1], size=256)
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param name: name of the gru group.
     :type name: basestring
     :param size: hidden size of the gru.
     :type size: int
-    :param reverse: whether to process the input data in a reverse order
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param act: type of the activiation
+    :param act: activiation type of gru
     :type act: BaseActivation
-    :param gate_act: type of the gate activiation
+    :param gate_act: gate activiation type of gru
     :type gate_act: BaseActivation
-    :param gru_bias_attr: bias. False means no bias, None means default bias.
-    :type gru_bias_attr: ParameterAttribute|False
-    :param gru_layer_attr: Extra parameter attribute of the gru layer.
-    :type gru_layer_attr: ParameterAttribute|False
+    :param gru_bias_attr: bias parameter attribute of gru layer, 
+                          False means no bias, None means default bias.
+    :type gru_bias_attr: ParameterAttribute|False|None
+    :param gru_layer_attr: Extra attribute of the gru layer.
+    :type gru_layer_attr: ExtraLayerAttribute
     :return: the gru group.
     :rtype: LayerOutput
     """
@@ -1145,7 +1152,7 @@ def bidirectional_gru(input,
                       concat_act=None):
     """
     A bidirectional_gru is a recurrent unit that iterates over the input
-    sequence both in forward and bardward orders, and then concatenate two
+    sequence both in forward and backward orders, and then concatenate two
     outputs to form a final output. However, concatenation of two outputs
     is not the only way to form the final output, you can also, for example,
     just add them together.
@@ -1162,11 +1169,10 @@ def bidirectional_gru(input,
     :type input: LayerOutput
     :param size: gru layer size.
     :type size: int
-    :param return_seq: If set False, outputs of the last time step are
-                       concatenated and returned.
-                       If set True, the entire output sequences that are
-                       processed in forward and backward directions are
+    :param return_seq: If set False, the last time step of output are
                        concatenated and returned.
+                       If set True, the entire output sequences in forward 
+                       and backward directions are concatenated and returned.
     :type return_seq: bool
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -1230,8 +1236,8 @@ def bidirectional_lstm(input,
                        concat_act=None):
     """
     A bidirectional_lstm is a recurrent unit that iterates over the input
-    sequence both in forward and bardward orders, and then concatenate two
-    outputs form a final output. However, concatenation of two outputs
+    sequence both in forward and backward orders, and then concatenate two
+    outputs to form a final output. However, concatenation of two outputs
     is not the only way to form the final output, you can also, for example,
     just add them together.
 
@@ -1252,13 +1258,12 @@ def bidirectional_lstm(input,
     :type input: LayerOutput
     :param size: lstm layer size.
     :type size: int
-    :param return_seq: If set False, outputs of the last time step are
-                       concatenated and returned.
-                       If set True, the entire output sequences that are
-                       processed in forward and backward directions are
+    :param return_seq: If set False, the last time step of output are
                        concatenated and returned.
+                       If set True, the entire output sequences in forward 
+                       and backward directions are concatenated and returned.
     :type return_seq: bool
-    :return: LayerOutput object accroding to the return_seq.
+    :return: LayerOutput object.
     :rtype: LayerOutput
     """
     args = locals()
@@ -1303,7 +1308,7 @@ def simple_attention(encoded_sequence,
                      weight_act=None,
                      name=None):
     """
-    Calculate and then return a context vector by attention machanism.
+    Calculate and return a context vector with attention mechanism.
     Size of the context vector equals to size of the encoded_sequence.
 
     ..  math::
@@ -1336,10 +1341,10 @@ def simple_attention(encoded_sequence,
     :param name: name of the attention model.
     :type name: basestring
     :param softmax_param_attr: parameter attribute of sequence softmax
-                               that is used to produce attention weight
+                               that is used to produce attention weight.
     :type softmax_param_attr: ParameterAttribute
-    :param weight_act: activation of the attention model
-    :type weight_act: Activation
+    :param weight_act: activation of the attention model.
+    :type weight_act: BaseActivation
     :param encoded_sequence: output of the encoder
     :type encoded_sequence: LayerOutput
     :param encoded_proj: attention weight is computed by a feed forward neural
@@ -1411,7 +1416,7 @@ def inputs(layers, *args):
 
 def outputs(layers, *args):
     """
-    Declare the outputs of network. If user have not defined the inputs of
+    Declare the outputs of network. If user has not defined the inputs of
     network, this method will calculate the input order by dfs travel.
 
     :param layers: Output layers.

python/paddle/v2/framework/op.py

@@ -215,5 +215,27 @@ class __RecurrentOp__(object):
         return core.RecurrentOp.create(proto.SerializeToString())
 
 
+class __CondOp__(object):
+    __proto__ = None
+    type = "cond"
+
+    def __init__(self):
+        # cache recurrent_op's proto
+        if self.__proto__ is None:
+            for op_proto in get_all_op_protos():
+                if op_proto.type == self.type:
+                    self.__proto__ = op_proto
+
+    def __call__(self, *args, **kwargs):
+        if self.type not in args and "type" not in kwargs:
+            kwargs["type"] = self.type
+        # create proto
+        create_method = OpDescCreationMethod(self.__proto__)
+        proto = create_method(*args, **kwargs)
+        # create condop
+        return core.CondOp.create(proto.SerializeToString())
+
+
 Operator = OperatorFactory()  # The default global factory
 RecurrentOp = __RecurrentOp__()
+CondOp = __CondOp__()

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

@@ -28,10 +28,10 @@ def create_op(scope, op_type, inputs, outputs, attrs):
         if out_name in outputs:
             kwargs[out_name] = []
             if out_dup:
-                sub_in = outputs[out_name]
-                for sub_in_name, _ in sub_in:
-                    var = scope.new_var(sub_in_name)
-                    kwargs[out_name].append(sub_in_name)
+                sub_out = outputs[out_name]
+                for sub_out_name, _ in sub_out:
+                    var = scope.new_var(sub_out_name)
+                    kwargs[out_name].append(sub_out_name)
             else:
                 var = scope.new_var(out_name)
                 kwargs[out_name].append(out_name)
@@ -39,6 +39,7 @@ def create_op(scope, op_type, inputs, outputs, attrs):
     for attr_name in Operator.get_op_attr_names(op_type):
         if attr_name in attrs:
             kwargs[attr_name] = attrs[attr_name]
+
     return Operator(op_type, **kwargs)
 
 
@@ -47,17 +48,24 @@ def set_input(scope, op, inputs, place):
         if in_name in inputs:
             if in_dup:
                 sub_in = inputs[in_name]
-                for sub_in_name, sub_in_array in sub_in:
+                for sub_in_name, sub_in_val in sub_in:
                     var = scope.find_var(sub_in_name)
                     tensor = var.get_tensor()
+                    sub_in_array = sub_in_val[0] \
+                        if isinstance(sub_in_val, tuple) else sub_in_val
                     tensor.set_dims(sub_in_array.shape)
                     tensor.set(sub_in_array, place)
+                    if isinstance(sub_in_val, tuple):
+                        tensor.set_lod(sub_in_val[1])
             else:
                 var = scope.find_var(in_name)
                 tensor = var.get_tensor()
-                arr = inputs[in_name]
-                tensor.set_dims(arr.shape)
-                tensor.set(arr, place)
+                in_val = inputs[in_name]
+                in_array = in_val[0] if isinstance(in_val, tuple) else in_val
+                tensor.set_dims(in_array.shape)
+                tensor.set(in_array, place)
+                if isinstance(in_val, tuple):
+                    tensor.set_lod(in_val[1])
 
 
 def set_output_grad(scope, op, outputs, place):
@@ -172,8 +180,9 @@ class OpTest(unittest.TestCase):
     def check_output_with_place(self, place):
         self.scope = core.Scope()
         op_inputs = self.inputs if hasattr(self, "inputs") else dict()
+        op_outputs = self.outputs if hasattr(self, "outputs") else dict()
         op_attrs = self.attrs if hasattr(self, "attrs") else dict()
-        self.op = create_op(self.scope, self.op_type, op_inputs, self.outputs,
+        self.op = create_op(self.scope, self.op_type, op_inputs, op_outputs,
                             op_attrs)
         if isinstance(place, core.GPUPlace) and not self.op.support_gpu():
             return
@@ -185,21 +194,26 @@ class OpTest(unittest.TestCase):
         for out_name, out_dup in Operator.get_op_outputs(self.op.type()):
             if out_dup:
                 sub_out = self.outputs[out_name]
-                for sub_out_name in sub_out:
+                if not isinstance(sub_out, list):
+                    raise AssertionError("sub_out type %s is not list",
+                                         type(sub_out))
+
+                for sub_out_name, expect in sub_out:
                     actual = np.array(
                         self.scope.find_var(sub_out_name).get_tensor())
-                    expect = sub_out[sub_out_name]
                     self.assertTrue(
                         np.allclose(
                             actual, expect, atol=1e-05),
-                        "output name: " + out_name + "has diff")
+                        "output name: " + out_name + " has diff")
             else:
-                actual = np.array(self.scope.find_var(out_name).get_tensor())
+                var = self.scope.find_var(out_name)
+                if var is not None:
+                    actual = np.array(var.get_tensor())
                     expect = self.outputs[out_name]
                     self.assertTrue(
                         np.allclose(
                             actual, expect, atol=1e-05),
-                    "output name: " + out_name + "has diff")
+                        "output name: " + out_name + " has diff")
 
     def check_output(self):
         places = [core.CPUPlace()]
@@ -234,8 +248,9 @@ class OpTest(unittest.TestCase):
                    max_relative_error=0.005):
         self.scope = core.Scope()
         op_inputs = self.inputs if hasattr(self, "inputs") else dict()
+        op_outputs = self.outputs if hasattr(self, "outputs") else dict()
         op_attrs = self.attrs if hasattr(self, "attrs") else dict()
-        self.op = create_op(self.scope, self.op_type, op_inputs, self.outputs,
+        self.op = create_op(self.scope, self.op_type, op_inputs, op_outputs,
                             op_attrs)
         if no_grad_set is None:
             no_grad_set = set()

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

@@ -6,16 +6,17 @@ from op_test import OpTest
 class TestAccuracyOp(OpTest):
     def setUp(self):
         self.op_type = "accuracy"
-        infer = np.random.randint(0, 2, (32, 1)).astype("int")
-        label = np.random.randint(0, 2, (32, )).astype("int")
+        n = 8192
+        infer = np.random.randint(0, 2, (n, 1)).astype("int")
+        label = np.random.randint(0, 2, (n, )).astype("int")
         self.inputs = {'Inference': infer, "Label": label}
         num_correct = 0
-        for rowid in xrange(32):
+        for rowid in xrange(n):
             for ele in infer[rowid]:
                 if ele == label[rowid]:
                     num_correct += 1
                     break
-        self.outputs = {'Accuracy': [num_correct / 32.0]}
+        self.outputs = {'Accuracy': [num_correct / float(n)]}
 
     def test_check_output(self):
         self.check_output()

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

+import logging
+import paddle.v2.framework.core as core
+import unittest
+import numpy as np
+from paddle.v2.framework.op import Operator, CondOp
+
+
+class PySimpleCond(object):
+    '''
+    A simple implementation of dynamic if-else based on numpy
+    '''
+
+    def __init__(self):
+        array = [1] * 10
+        for i in range(1, 10, 2):
+            array[i] = 0
+        self.cond = np.array(array)
+        self.x = np.ones(shape=(10, 1))
+
+    def forward(self):
+        self.index_t = np.where(self.cond == 1)
+        self.index_f = np.where(self.cond == 0)
+        y_t = self.x[self.index_t]
+        y_f = self.x[self.index_f]
+        y_t = y_t * 2.
+        y_f = y_f * (-2.)
+        output = np.zeros(shape=(10, 1))
+        output[self.index_t] = y_t
+        output[self.index_f] = y_f
+        return output
+
+
+class PySimpleCondTest(unittest.TestCase):
+    def setUp(self):
+        self.condnn = PySimpleCond()
+
+    def test_forward(self):
+        output = self.condnn.forward()
+
+
+def create_tensor(scope, name, shape, np_data):
+    tensor = scope.new_var(name).get_tensor()
+    tensor.set_dims(shape)
+    tensor.set(np_data, core.CPUPlace())
+    return tensor
+
+
+class TestCondOp(unittest.TestCase):
+    '''
+    Test CondOp
+
+    equation:
+        cond = [True, False, True, False, ...]
+        y[index_t] = x[index_t] * 2.
+        y[index_f] = x[index_f] * -2.
+    outputs:
+        y
+    '''
+
+    def setUp(self):
+        self.py_cond = PySimpleCond()
+
+    def forward(self):
+        self.scope = core.Scope()
+        self.create_global_variables()
+        self.create_cond_op()
+        self.create_sub_net()
+        ctx = core.DeviceContext.create(core.CPUPlace())
+        self.condop.infer_shape(self.scope)
+        self.condop.run(self.scope, ctx)
+        return np.array(self.scope.find_var("Out").get_tensor())
+
+    def create_global_variables(self):
+        x_np_data = self.py_cond.x
+        create_tensor(self.scope, "X", [10, 1], x_np_data)
+        cond_np_data = self.py_cond.cond.astype("int32")
+        create_tensor(self.scope, "cond", [10, 1], cond_np_data)
+        self.scope.new_var("SubScopes")
+        self.scope.new_var("IndexTensors")
+        self.scope.new_var("Out")
+
+    def create_cond_op(self):
+        self.condop = CondOp(
+            Cond="cond",
+            Xs=["X"],
+            Outs=["Out"],
+            SubScopes="SubScopes",
+            IndexTensors="IndexTensors")
+
+    def create_sub_net(self):
+        truenet = core.Net.create()
+        scale_op_t = Operator("scale", X='X', Out='Out', scale=2.)
+        truenet.append_op(scale_op_t)
+        truenet.complete_add_op(True)
+        self.condop.set_truenet(truenet)
+
+        falsenet = core.Net.create()
+        scale_op_t = Operator("scale", X='X', Out='Out', scale=-2.)
+        falsenet.append_op(scale_op_t)
+        falsenet.complete_add_op(True)
+        self.condop.set_falsenet(falsenet)
+
+    def test_forward(self):
+        print 'test cond op forward'
+        pd_output = self.forward()
+        py_output = self.py_cond.forward()
+        print 'pd_output', pd_output
+        print
+        print 'py_output', py_output
+        self.assertEqual(pd_output.shape, py_output.shape)
+        print 'test passed'
+        return 0
+
+
+if __name__ == "__main__":
+    unittest.main()

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

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestFCOp1(OpTest):
+    def setUp(self):
+        x0 = np.random.random((16, 32)).astype("float32")
+        w0 = np.random.random((32, 10)).astype("float32")
+
+        mul_out0 = np.dot(x0, w0)
+        identity_out = mul_out0
+
+        self.op_type = "fc"
+        self.inputs = {"X": [("X0", x0)], "W": [("W0", w0)]}
+        self.outputs = {"MulOut": [("MulOut0", mul_out0)], "Out": identity_out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X0", "W0"], "Out", max_relative_error=0.01)
+
+
+class TestFCOp2(OpTest):
+    def setUp(self):
+        x0 = np.random.random((16, 4, 8)).astype("float32")
+        x1 = np.random.random((4, 4, 32)).astype("float32")
+        w0 = np.random.random((32, 10)).astype("float32")
+        w1 = np.random.random((32, 10)).astype("float32")
+        b = np.random.random(10).astype("float32")
+
+        mul_out0 = np.dot(x0.reshape(16, 4 * 8), w0)
+        mul_out1 = np.dot(x1.reshape(4 * 4, 32), w1)
+        sum_out = mul_out0 + mul_out1
+        add_out = np.add(sum_out, b)
+        sigmoid_out = 1 / (1 + np.exp(-add_out))
+
+        self.op_type = "fc"
+        self.inputs = {
+            "X": [("X0", x0), ("X1", x1)],
+            "W": [("W0", w0), ("W1", w1)],
+            "B": b
+        }
+        self.attrs = {"xNumColDims": [1, 2], "activation": "sigmoid"}
+        self.outputs = {
+            "MulOut": [("MulOut0", mul_out0), ("MulOut1", mul_out1)],
+            "SumOut": sum_out,
+            "AddOut": add_out,
+            "Out": sigmoid_out
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(
+            ["X0", "X1", "W0", "W1", "B"], "Out", max_relative_error=0.01)
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -4,7 +4,7 @@ from paddle.v2.framework.op import Operator
 import numpy
 
 
-class GaussianRandomTest(unittest.TestCase):
+class TestGaussianRandomOp(unittest.TestCase):
     def test_cpu(self):
         self.gaussian_random_test(place=core.CPUPlace())
 

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

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestIdentityOp(OpTest):
+    def setUp(self):
+        self.op_type = "identity"
+        self.inputs = {'X': np.random.random((10, 10)).astype("float32")}
+        self.outputs = {'Y': self.inputs['X']}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Y')
+
+
+if __name__ == "__main__":
+    unittest.main()

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

@@ -3,7 +3,7 @@ import numpy as np
 from op_test import OpTest
 
 
-class MinusOpTest(OpTest):
+class TestMinusOp(OpTest):
     def setUp(self):
         self.op_type = "minus"
         self.inputs = {

python/paddle/v2/framework/tests/test_cross_entropy_op.py → python/paddle/v2/framework/tests/test_onehot_cross_entropy_op.py

@@ -3,7 +3,7 @@ import numpy
 from op_test import OpTest
 
 
-class TestCrossEntropy(OpTest):
+class TestOnehotCrossEntropyOp(OpTest):
     def setUp(self):
         self.op_type = "onehot_cross_entropy"
         batch_size = 30

python/paddle/v2/framework/tests/test_scale_and_identity_op.py → python/paddle/v2/framework/tests/test_scale_op.py

@@ -3,20 +3,7 @@ import numpy as np
 from op_test import OpTest
 
 
-class IdentityTest(OpTest):
-    def setUp(self):
-        self.op_type = "identity"
-        self.inputs = {'X': np.random.random((10, 10)).astype("float32")}
-        self.outputs = {'Out': self.inputs['X']}
-
-    def test_check_output(self):
-        self.check_output()
-
-    def test_check_grad(self):
-        self.check_grad(['X'], 'Out')
-
-
-class ScaleTest(OpTest):
+class TestScaleOp(OpTest):
     def setUp(self):
         self.op_type = "scale"
         self.inputs = {'X': np.random.random((10, 10)).astype("float32")}

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

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestSeqAvgPool1D(OpTest):
+    def setUp(self):
+        self.op_type = 'sequence_avg_pool'
+        # one level, batch size is 4
+        x = np.random.uniform(0.1, 1, [11, 23]).astype('float32')
+        lod = [[0, 4, 5, 8, 11]]
+
+        out = np.zeros((4, 23)).astype('float32')
+        for i in range(4):
+            sub_x = x[lod[0][i]:lod[0][i + 1], :]
+            out[i] = sub_x.mean(axis=0)
+
+        self.inputs = {'X': (x, lod)}
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+
+class TestSeqAvgPool2D(OpTest):
+    def setUp(self):
+        self.op_type = 'sequence_avg_pool'
+        # one level, batch size is 4
+        x = np.random.uniform(0.1, 1, [13, 3, 17]).astype('float32')
+        lod = [[0, 4, 5, 8, 13]]
+
+        out = np.zeros((4, 3, 17)).astype('float32')
+        for i in range(4):
+            sub_x = np.reshape(x[lod[0][i]:lod[0][i + 1], :], (-1, 3 * 17))
+            out[i] = np.reshape(sub_x.mean(axis=0), (3, 17))
+
+        self.inputs = {'X': (x, lod)}
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -3,7 +3,7 @@ import numpy as np
 from op_test import OpTest
 
 
-class TestSGD(OpTest):
+class TestSGDOp(OpTest):
     def setUp(self):
         self.op_type = "sgd"
         w = np.random.random((102, 105)).astype("float32")

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

@@ -3,7 +3,7 @@ import numpy as np
 from op_test import OpTest
 
 
-class TestSigmoid(OpTest):
+class TestSigmoidOp(OpTest):
     def setUp(self):
         self.op_type = "sigmoid"
         self.inputs = {

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

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestSplitOp(OpTest):
+    def setUp(self):
+        self.op_type = "split"
+        axis = 0
+        num = 2
+        x = np.random.random((4, 2)).astype('float32')
+        out = np.split(x, num, axis)
+        self.inputs = {'X': x}
+        self.attrs = {'axis': axis, 'num': num}
+        self.outputs = {'Out': [('out%d' % i, out[i]) \
+            for i in xrange(len(out))]}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], ['out0', 'out1'])
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -21,6 +21,9 @@ class TestTopkOp(OpTest):
 
         self.outputs = {'Out': output, 'Indices': indices}
 
+    def test_check_output(self):
+        self.check_output()
+
 
 class TestTopkOp3d(OpTest):
     def setUp(self):
@@ -42,6 +45,9 @@ class TestTopkOp3d(OpTest):
 
         self.outputs = {'Out': output, 'Indices': indices}
 
+    def test_check_output(self):
+        self.check_output()
+
 
 if __name__ == "__main__":
     unittest.main()

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

@@ -4,7 +4,7 @@ import paddle.v2.framework.core as core
 import numpy
 
 
-class UniformRandomTest(unittest.TestCase):
+class TestUniformRandomOp(unittest.TestCase):
     def test_uniform_random_cpu(self):
         self.uniform_random_test(place=core.CPUPlace())