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

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);
+}
+```

paddle/cuda/include/hl_cuda_cudnn.h

@@ -22,10 +22,10 @@ limitations under the License. */
  */
 typedef enum {
   HL_POOLING_MAX = 0,
-  // average includes padded values
-  HL_POOLING_AVERAGE = 1,
   // average does not include padded values
-  HL_POOLING_AVERAGE_EXCLUDE_PADDING = 2,
+  HL_POOLING_AVERAGE = 1,
+  // average includes padded values
+  HL_POOLING_AVERAGE_INCLUDE_PADDING = 2,
   HL_POOLING_END
 } hl_pooling_mode_t;
 

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/cuda/src/hl_cuda_cnn.cu

@@ -211,13 +211,11 @@ __global__ void KeAvgPoolForward(const int nthreads,
 
     int hstart = ph * strideH - padH;
     int wstart = pw * strideW - padW;
-    int hend = min(hstart + sizeY, height + padH);
-    int wend = min(wstart + sizeX, width + padW);
-    int pool_size = (hend - hstart) * (wend - wstart);
+    int hend = min(hstart + sizeY, height);
+    int wend = min(wstart + sizeX, width);
     hstart = max(hstart, 0);
     wstart = max(wstart, 0);
-    hend = min(hend, height);
-    wend = min(wend, width);
+    int pool_size = (hend - hstart) * (wend - wstart);
 
     real aveval = 0;
     inputData += (frameNum * channels + c) * height * width;
@@ -299,12 +297,14 @@ __global__ void KeAvgPoolBackward(const int nthreads,
     outGrad += (frameNum * outStride + offsetC * pooledH * pooledW);
 
     for (int ph = phstart; ph < phend; ++ph) {
+      int hstart = ph * strideH - padH;
+      int hend = min(hstart + sizeY, height);
+      hstart = max(hstart, 0);
       for (int pw = pwstart; pw < pwend; ++pw) {
         // figure out the pooling size
-        int hstart = ph * strideH - padH;
         int wstart = pw * strideW - padW;
-        int hend = min(hstart + sizeY, height + padH);
-        int wend = min(wstart + sizeX, width + padW);
+        int wend = min(wstart + sizeX, width);
+        wstart = max(wstart, 0);
         int poolsize = (hend - hstart) * (wend - wstart);
         gradient += outGrad[ph * pooledW + pw] / poolsize;
       }
@@ -600,16 +600,13 @@ __global__ void KeAvgPool3DForward(const int nthreads,
     int dstart = pd * strideD - padD;
     int hstart = ph * strideH - padH;
     int wstart = pw * strideW - padW;
-    int dend = min(dstart + sizeZ, depth + padD);
-    int hend = min(hstart + sizeY, height + padH);
-    int wend = min(wstart + sizeX, width + padW);
-    int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart);
+    int dend = min(dstart + sizeZ, depth);
+    int hend = min(hstart + sizeY, height);
+    int wend = min(wstart + sizeX, width);
     dstart = max(dstart, 0);
     hstart = max(hstart, 0);
     wstart = max(wstart, 0);
-    dend = min(dend, depth);
-    hend = min(hend, height);
-    wend = min(wend, width);
+    int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart);
 
     real aveval = 0;
     inputData += (frameNum * channels + c) * depth * height * width;
@@ -712,15 +709,18 @@ __global__ void KeAvgPool3DBackward(const int nthreads,
     outGrad += (frameNum * channels + offsetC) * pooledD * pooledH * pooledW;
 
     for (int pd = pdstart; pd < pdend; ++pd) {
+      int dstart = pd * strideD - padD;
+      int dend = min(dstart + sizeZ, depth);
+      dstart = max(dstart, 0);
       for (int ph = phstart; ph < phend; ++ph) {
+        int hstart = ph * strideH - padH;
+        int hend = min(hstart + sizeY, height);
+        hstart = max(hstart, 0);
         for (int pw = pwstart; pw < pwend; ++pw) {
           // figure out the pooling size
-          int dstart = pd * strideD - padD;
-          int hstart = ph * strideH - padH;
           int wstart = pw * strideW - padW;
-          int dend = min(dstart + sizeZ, depth + padD);
-          int hend = min(hstart + sizeY, height + padH);
-          int wend = min(wstart + sizeX, width + padW);
+          int wend = min(wstart + sizeX, width);
+          wstart = max(wstart, 0);
           int poolsize = (dend - dstart) * (hend - hstart) * (wend - wstart);
           gradient += outGrad[(pd * pooledH + ph) * pooledW + pw] / poolsize;
         }

paddle/cuda/src/hl_cuda_cudnn.cc

@@ -432,11 +432,11 @@ void hl_create_pooling_descriptor(hl_pooling_descriptor* pooling_desc,
       cudnn_mode = CUDNN_POOLING_MAX;
       break;
     case HL_POOLING_AVERAGE:
-      cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING;
-      break;
-    case HL_POOLING_AVERAGE_EXCLUDE_PADDING:
       cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING;
       break;
+    case HL_POOLING_AVERAGE_INCLUDE_PADDING:
+      cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING;
+      break;
     default:
       LOG(FATAL) << "parameter mode error";
   }

paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp

@@ -18,7 +18,6 @@ limitations under the License. */
 #include <cmath>
 #include <functional>
 #include <limits>
-#include <memory>
 #include "NeuralNetwork.h"
 #include "paddle/gserver/layers/AgentLayer.h"
 #include "paddle/utils/Flags.h"
@@ -430,11 +429,7 @@ void RecurrentGradientMachine::reorganizeInput(PassType passType) {
   }
 
   {
-    std::unique_ptr<AsyncGpuBlock> asyncBlock;
-
-    if (useGpu_) {
-      asyncBlock.reset(new AsyncGpuBlock());
-    }
+    AsyncGpuBlock asyncGpuBlock;
 
     // inFrameLine select rows in real layer one time
     for (size_t i = 0; i < inFrameLines_.size(); i++) {

paddle/gserver/layers/CudnnPoolLayer.cpp

@@ -29,9 +29,9 @@ bool CudnnPoolLayer::typeCheck(const std::string &poolType,
     if (mode) {
       *mode = HL_POOLING_AVERAGE;
     }
-  } else if (poolType == "cudnn-avg-excl-pad-pool") {
+  } else if (poolType == "cudnn-avg-incl-pad-pool") {
     if (mode) {
-      *mode = HL_POOLING_AVERAGE_EXCLUDE_PADDING;
+      *mode = HL_POOLING_AVERAGE_INCLUDE_PADDING;
     }
   } else {
     return false;

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/Matrix.cpp

@@ -1033,17 +1033,15 @@ void GpuMatrix::maxPoolForward(Matrix& inputMat,
 
   real* inputData = inputMat.getData();
   size_t frameNum = inputMat.getHeight();
-  size_t width = imgSizeW;
-  size_t height = imgSizeH;
-  CHECK(height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
   CHECK(width_ == outputH * outputW * channels);
 
   hl_maxpool_forward(frameNum,
                      inputData,
                      channels,
-                     height,
-                     width,
+                     imgSizeH,
+                     imgSizeW,
                      outputH,
                      outputW,
                      sizeX,
@@ -1080,11 +1078,8 @@ void GpuMatrix::maxPoolBackward(Matrix& inputMat,
   real* outDiff = outGrad.getData();
   size_t frameNum = inputMat.getHeight();
   size_t channels = outV.getWidth() / outputH / outputW;
-  size_t width = imgSizeW;
-  size_t height = imgSizeH;
-  CHECK(height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
-  CHECK(width_ == width * height * channels);
   CHECK(outGrad.getHeight() == outV.getHeight() &&
         outGrad.getWidth() == outV.getWidth());
 
@@ -1093,8 +1088,8 @@ void GpuMatrix::maxPoolBackward(Matrix& inputMat,
                       outData,
                       outDiff,
                       channels,
-                      height,
-                      width,
+                      imgSizeH,
+                      imgSizeW,
                       outputH,
                       outputW,
                       sizeX,
@@ -1125,17 +1120,15 @@ void GpuMatrix::avgPoolForward(Matrix& inputMat,
 
   real* inputData = inputMat.getData();
   size_t frameNum = inputMat.getHeight();
-  size_t height = imgSizeH;
-  size_t width = imgSizeW;
-  CHECK(height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
   CHECK(width_ == outputH * outputW * channels);
 
   hl_avgpool_forward(frameNum,
                      inputData,
                      channels,
-                     height,
-                     width,
+                     imgSizeH,
+                     imgSizeW,
                      outputH,
                      outputW,
                      sizeX,
@@ -1166,17 +1159,15 @@ void GpuMatrix::avgPoolBackward(Matrix& outGrad,
   real* outDiff = outGrad.getData();
   size_t frameNum = outGrad.getHeight();
   size_t channels = outGrad.getWidth() / outputH / outputW;
-  size_t height = imgSizeH;
-  size_t width = imgSizeW;
-  CHECK(height * width * channels == width_);
+  CHECK(imgSizeH * imgSizeW * channels == width_);
   CHECK(height_ == outGrad.getHeight());
   CHECK(outGrad.getWidth() == outputH * outputW * channels);
 
   hl_avgpool_backward(frameNum,
                       outDiff,
                       channels,
-                      height,
-                      width,
+                      imgSizeH,
+                      imgSizeW,
                       outputH,
                       outputW,
                       sizeX,
@@ -1214,19 +1205,16 @@ void GpuMatrix::maxPool3DForward(Matrix& inputMat,
   real* inputData = inputMat.getData();
   real* maxPoolIdxData = maxPoolIdx.getData();
   size_t num = inputMat.getHeight();
-  size_t width = imgSizeW;
-  size_t height = imgSizeH;
-  size_t depth = imgSizeD;
-  CHECK(depth * height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
   CHECK(width_ == outputD * outputH * outputW * channels);
 
   hl_maxpool3D_forward(num,
                        inputData,
                        channels,
-                       depth,
-                       height,
-                       width,
+                       imgSizeD,
+                       imgSizeH,
+                       imgSizeW,
                        outputD,
                        outputH,
                        outputW,
@@ -1269,20 +1257,16 @@ void GpuMatrix::maxPool3DBackward(Matrix& outGrad,
   real* maxPoolIdxData = maxPoolIdx.getData();
   size_t frameNum = getHeight();
   size_t channels = outGrad.getWidth() / outputD / outputH / outputW;
-  size_t width = imgSizeW;
-  size_t height = imgSizeH;
-  size_t depth = imgSizeD;
-  CHECK(depth * height * width * channels == getWidth());
-  CHECK(width_ == depth * width * height * channels);
+  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth());
   CHECK(outGrad.getHeight() == maxPoolIdx.getHeight() &&
         outGrad.getWidth() == maxPoolIdx.getWidth());
 
   hl_maxpool3D_backward(frameNum,
                         outDiff,
                         channels,
-                        depth,
-                        height,
-                        width,
+                        imgSizeD,
+                        imgSizeH,
+                        imgSizeW,
                         outputD,
                         outputH,
                         outputW,
@@ -1323,19 +1307,16 @@ void GpuMatrix::avgPool3DForward(Matrix& inputMat,
 
   real* inputData = inputMat.getData();
   size_t frameNum = inputMat.getHeight();
-  size_t height = imgSizeH;
-  size_t width = imgSizeW;
-  size_t depth = imgSizeD;
-  CHECK(depth * height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
   CHECK(width_ == outputD * outputH * outputW * channels);
 
   hl_avgpool3D_forward(frameNum,
                        inputData,
                        channels,
-                       depth,
-                       height,
-                       width,
+                       imgSizeD,
+                       imgSizeH,
+                       imgSizeW,
                        outputD,
                        outputH,
                        outputW,
@@ -1375,19 +1356,16 @@ void GpuMatrix::avgPool3DBackward(Matrix& outGrad,
   real* outDiff = outGrad.getData();
   size_t frameNum = outGrad.getHeight();
   size_t channels = outGrad.getWidth() / outputD / outputH / outputW;
-  size_t height = imgSizeH;
-  size_t width = imgSizeW;
-  size_t depth = imgSizeD;
-  CHECK(depth * height * width * channels == width_);
+  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == width_);
   CHECK(height_ == outGrad.getHeight());
   CHECK(outGrad.getWidth() == outputD * outputH * outputW * channels);
 
   hl_avgpool3D_backward(frameNum,
                         outDiff,
                         channels,
-                        depth,
-                        height,
-                        width,
+                        imgSizeD,
+                        imgSizeH,
+                        imgSizeW,
                         outputD,
                         outputH,
                         outputW,
@@ -1999,11 +1977,11 @@ void CpuMatrix::maxPoolForward(Matrix& inputMat,
   real* inputData = inputMat.getData();
   real* outData = data_;
   size_t num = inputMat.getHeight();
-  size_t inWidth = imgSizeW;
-  size_t inHeight = imgSizeH;
-  CHECK(inHeight * inWidth == inputMat.getWidth() / channels);
+  size_t inLength = imgSizeH * imgSizeW;
+  size_t outLength = outputH * outputW;
+  CHECK(inLength == inputMat.getWidth() / channels);
   CHECK_EQ(num, this->getHeight());
-  CHECK_EQ(channels * outputH * outputW, this->getWidth());
+  CHECK_EQ(channels * outLength, this->getWidth());
   size_t outStride = getStride();
 
   /* initialize the data_ */
@@ -2020,24 +1998,24 @@ void CpuMatrix::maxPoolForward(Matrix& inputMat,
     }
     for (size_t c = 0; c < channels; ++c) {  // channel by channel
       for (size_t ph = 0; ph < outputH; ++ph) {
-        for (size_t pw = 0; pw < outputW; ++pw) {
         int hstart = ph * strideH - paddingH;
-          int wstart = pw * strideW - paddingW;
-          int hend = std::min(hstart + sizeY, inHeight);
-          int wend = std::min(wstart + sizeX, inWidth);
+        int hend = std::min(hstart + sizeY, imgSizeH);
         hstart = std::max(hstart, 0);
+        for (size_t pw = 0; pw < outputW; ++pw) {
+          int wstart = pw * strideW - paddingW;
+          int wend = std::min(wstart + sizeX, imgSizeW);
           wstart = std::max(wstart, 0);
           for (int h = hstart; h < hend; ++h) {
             for (int w = wstart; w < wend; ++w) {
-              outData[ph * outputW + pw] = std::max(outData[ph * outputW + pw],
-                                                    inputData[h * inWidth + w]);
+              outData[ph * outputW + pw] = std::max(
+                  outData[ph * outputW + pw], inputData[h * imgSizeW + w]);
             }
           }
         }
       }
       // compute offset
-      inputData += inHeight * inWidth;
-      outData += outputH * outputW;
+      inputData += inLength;
+      outData += outLength;
     }
   }
 }
@@ -2058,8 +2036,10 @@ void CpuMatrix::maxPoolBackward(Matrix& image,
                                 size_t paddingH,
                                 size_t paddingW) {
   size_t num = image.getHeight();
-  size_t channels = size_t(width_ / imgSizeH / imgSizeW);
-  CHECK(image.getWidth() == imgSizeH * imgSizeW * channels);
+  size_t inLength = imgSizeH * imgSizeW;
+  size_t outLength = outputH * outputW;
+  size_t channels = size_t(width_ / inLength);
+  CHECK(image.getWidth() == inLength * channels);
   CHECK(image.getHeight() == height_ && image.getWidth() == width_);
   CHECK(outV.getHeight() == outGrad.getHeight() &&
         outV.getWidth() == outGrad.getWidth());
@@ -2080,12 +2060,12 @@ void CpuMatrix::maxPoolBackward(Matrix& image,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t ph = 0; ph < outputH; ++ph) {
-        for (size_t pw = 0; pw < outputW; ++pw) {
         int hstart = ph * strideH - paddingH;
-          int wstart = pw * strideW - paddingW;
         int hend = std::min(hstart + sizeY, imgSizeH);
-          int wend = std::min(wstart + sizeX, imgSizeW);
         hstart = std::max(hstart, 0);
+        for (size_t pw = 0; pw < outputW; ++pw) {
+          int wstart = pw * strideW - paddingW;
+          int wend = std::min(wstart + sizeX, imgSizeW);
           wstart = std::max(wstart, 0);
           for (int h = hstart; h < hend; ++h) {
             for (int w = wstart; w < wend; ++w) {
@@ -2098,10 +2078,10 @@ void CpuMatrix::maxPoolBackward(Matrix& image,
         }
       }
       // offset
-      inData += imgSizeH * imgSizeW;
-      tgtGrad += imgSizeH * imgSizeW;
-      otData += outputH * outputW;
-      otGrad += outputH * outputW;
+      inData += inLength;
+      tgtGrad += inLength;
+      otData += outLength;
+      otGrad += outLength;
     }
   }
 }
@@ -2120,10 +2100,10 @@ void CpuMatrix::avgPoolForward(Matrix& input,
                                size_t paddingW) {
   // The main loop
   size_t num = input.getHeight();
-  size_t inHeight = imgSizeH;
-  size_t inWidth = imgSizeW;
-  CHECK(inHeight * inWidth * channels == input.getWidth());
-  CHECK(outputH * outputW * channels * num == height_ * width_);
+  size_t inLength = imgSizeH * imgSizeW;
+  size_t outLength = outputH * outputW;
+  CHECK(inLength * channels == input.getWidth());
+  CHECK(outLength * channels * num == height_ * width_);
   real* tgtData = data_;
   real* inData = input.getData();
 
@@ -2133,30 +2113,27 @@ void CpuMatrix::avgPoolForward(Matrix& input,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t ph = 0; ph < outputH; ++ph) {
-        for (size_t pw = 0; pw < outputW; ++pw) {
         int hstart = ph * strideH - paddingH;
-          int wstart = pw * strideW - paddingW;
-          int hend = std::min(hstart + sizeY, inHeight + paddingH);
-          int wend = std::min(wstart + sizeX, inWidth + paddingW);
-          int poolSize = (hend - hstart) * (wend - wstart);
+        int hend = std::min(hstart + sizeY, imgSizeH);
         hstart = std::max(hstart, 0);
+        for (size_t pw = 0; pw < outputW; ++pw) {
+          int wstart = pw * strideW - paddingW;
+          int wend = std::min(wstart + sizeX, imgSizeW);
           wstart = std::max(wstart, 0);
-          hend = std::min(hend, static_cast<int>(inHeight));
-          wend = std::min(wend, static_cast<int>(inWidth));
-
-          CHECK(poolSize);
           tgtData[ph * outputW + pw] = 0;  // clear
           for (int h = hstart; h < hend; ++h) {
             for (int w = wstart; w < wend; ++w) {
-              tgtData[ph * outputW + pw] += inData[h * inWidth + w];
+              tgtData[ph * outputW + pw] += inData[h * imgSizeW + w];
             }
           }
+          int poolSize = (hend - hstart) * (wend - wstart);
+          CHECK(poolSize);
           tgtData[ph * outputW + pw] /= poolSize;
         }
       }
       // compute offset
-      inData += inHeight * inWidth;
-      tgtData += outputH * outputW;
+      inData += inLength;
+      tgtData += outLength;
     }
   }
 }
@@ -2176,7 +2153,9 @@ void CpuMatrix::avgPoolBackward(Matrix& input,
                                 size_t paddingW) {
   size_t num = input.getHeight();
   size_t channels = input.getWidth() / outputH / outputW;
-  CHECK(imgSizeH * imgSizeW * channels == getWidth());
+  size_t inLength = imgSizeH * imgSizeW;
+  size_t outLength = outputH * outputW;
+  CHECK(inLength * channels == getWidth());
   real* inData = input.getData();
   real* outData = getData();
 
@@ -2186,16 +2165,14 @@ void CpuMatrix::avgPoolBackward(Matrix& input,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t ph = 0; ph < outputH; ++ph) {
-        for (size_t pw = 0; pw < outputW; ++pw) {
         int hstart = ph * strideH - paddingH;
-          int wstart = pw * strideW - paddingW;
-          int hend = std::min(hstart + sizeY, imgSizeH + paddingH);
-          int wend = std::min(wstart + sizeX, imgSizeW + paddingW);
-          int poolSize = (hend - hstart) * (wend - wstart);
+        int hend = std::min(hstart + sizeY, imgSizeH);
         hstart = std::max(hstart, 0);
+        for (size_t pw = 0; pw < outputW; ++pw) {
+          int wstart = pw * strideW - paddingW;
+          int wend = std::min(wstart + sizeX, imgSizeW);
           wstart = std::max(wstart, 0);
-          hend = std::min(hend, static_cast<int>(imgSizeH));
-          wend = std::min(wend, static_cast<int>(imgSizeW));
+          int poolSize = (hend - hstart) * (wend - wstart);
           CHECK(poolSize);
 
           for (int h = hstart; h < hend; ++h) {
@@ -2206,8 +2183,8 @@ void CpuMatrix::avgPoolBackward(Matrix& input,
         }
       }
       // offset
-      outData += imgSizeH * imgSizeW;
-      inData += outputH * outputW;
+      outData += inLength;
+      inData += outLength;
     }
   }
 }
@@ -2234,12 +2211,11 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat,
   real* outData = getData();
   real* maxPoolIdxData = maxPoolIdx.getData();
   size_t num = inputMat.getHeight();
-  size_t inWidth = imgSizeW;
-  size_t inHeight = imgSizeH;
-  size_t inDepth = imgSizeD;
-  CHECK(inHeight * inWidth * inDepth == inputMat.getWidth() / channels);
+  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
+  size_t outLength = outputH * outputW * outputD;
+  CHECK(inLength == inputMat.getWidth() / channels);
   CHECK_EQ(num, this->getHeight());
-  CHECK_EQ(channels * outputH * outputW * outputD, this->getWidth());
+  CHECK_EQ(channels * outLength, this->getWidth());
   size_t outStride = getStride();
 
   /* initialize the data_ */
@@ -2258,16 +2234,16 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat,
     }
     for (size_t c = 0; c < channels; ++c) {  // channel by channel
       for (size_t pd = 0; pd < outputD; ++pd) {
-        for (size_t ph = 0; ph < outputH; ++ph) {
-          for (size_t pw = 0; pw < outputW; ++pw) {
         int dstart = pd * strideD - paddingD;
-            int hstart = ph * strideH - paddingH;
-            int wstart = pw * strideW - paddingW;
-            int dend = std::min(dstart + sizeZ, inDepth);
-            int hend = std::min(hstart + sizeY, inHeight);
-            int wend = std::min(wstart + sizeX, inWidth);
+        int dend = std::min(dstart + sizeZ, imgSizeD);
         dstart = std::max(dstart, 0);
+        for (size_t ph = 0; ph < outputH; ++ph) {
+          int hstart = ph * strideH - paddingH;
+          int hend = std::min(hstart + sizeY, imgSizeH);
           hstart = std::max(hstart, 0);
+          for (size_t pw = 0; pw < outputW; ++pw) {
+            int wstart = pw * strideW - paddingW;
+            int wend = std::min(wstart + sizeX, imgSizeW);
             wstart = std::max(wstart, 0);
             int maxIdx = -1;
             real maxOutData = outData[(pd * outputH + ph) * outputW + pw];
@@ -2275,9 +2251,9 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat,
               for (int h = hstart; h < hend; ++h) {
                 for (int w = wstart; w < wend; ++w) {
                   if (maxOutData <
-                      inputData[(d * inHeight + h) * inWidth + w]) {
-                    maxOutData = inputData[(d * inHeight + h) * inWidth + w];
-                    maxIdx = (d * inHeight + h) * inWidth + w;
+                      inputData[(d * imgSizeH + h) * imgSizeW + w]) {
+                    maxOutData = inputData[(d * imgSizeH + h) * imgSizeW + w];
+                    maxIdx = (d * imgSizeH + h) * imgSizeW + w;
                   }
                 }
               }
@@ -2288,9 +2264,9 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat,
         }
       }
       // compute offset
-      inputData += inDepth * inHeight * inWidth;
-      outData += outputD * outputH * outputW;
-      maxPoolIdxData += outputD * outputH * outputW;
+      inputData += inLength;
+      outData += outLength;
+      maxPoolIdxData += outLength;
     }
   }
 }
@@ -2315,7 +2291,9 @@ void CpuMatrix::maxPool3DBackward(Matrix& outGrad,
                                   real scaleTargets,
                                   real scaleOutput) {
   size_t num = getHeight();
-  size_t channels = size_t(width_ / imgSizeD / imgSizeH / imgSizeW);
+  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
+  size_t outLength = outputH * outputW * outputD;
+  size_t channels = size_t(width_ / inLength);
   CHECK(maxPoolIdx.getHeight() == outGrad.getHeight() &&
         maxPoolIdx.getWidth() == outGrad.getWidth());
 
@@ -2341,9 +2319,9 @@ void CpuMatrix::maxPool3DBackward(Matrix& outGrad,
         }
       }
       // offset
-      tgtGrad += imgSizeD * imgSizeH * imgSizeW;
-      otGrad += outputD * outputH * outputW;
-      maxPoolIdxData += outputD * outputH * outputW;
+      tgtGrad += inLength;
+      otGrad += outLength;
+      maxPoolIdxData += outLength;
     }
   }
 }
@@ -2367,11 +2345,10 @@ void CpuMatrix::avgPool3DForward(Matrix& input,
                                  size_t paddingW) {
   // The main loop
   size_t num = input.getHeight();
-  size_t inDepth = imgSizeD;
-  size_t inHeight = imgSizeH;
-  size_t inWidth = imgSizeW;
-  CHECK(inDepth * inHeight * inWidth * channels == input.getWidth());
-  CHECK(outputD * outputH * outputW * channels * num == height_ * width_);
+  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
+  size_t outLength = outputH * outputW * outputD;
+  CHECK(inLength * channels == input.getWidth());
+  CHECK(outLength * channels * num == height_ * width_);
   real* tgtData = getData();
   real* inData = input.getData();
 
@@ -2381,39 +2358,36 @@ void CpuMatrix::avgPool3DForward(Matrix& input,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t pd = 0; pd < outputD; ++pd) {
-        for (size_t ph = 0; ph < outputH; ++ph) {
-          for (size_t pw = 0; pw < outputW; ++pw) {
         int dstart = pd * strideD - paddingD;
-            int hstart = ph * strideH - paddingH;
-            int wstart = pw * strideW - paddingW;
-            int dend = std::min(dstart + sizeZ, inDepth + paddingD);
-            int hend = std::min(hstart + sizeY, inHeight + paddingH);
-            int wend = std::min(wstart + sizeX, inWidth + paddingW);
-            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
+        int dend = std::min(dstart + sizeZ, imgSizeD);
         dstart = std::max(dstart, 0);
+        for (size_t ph = 0; ph < outputH; ++ph) {
+          int hstart = ph * strideH - paddingH;
+          int hend = std::min(hstart + sizeY, imgSizeH);
           hstart = std::max(hstart, 0);
+          for (size_t pw = 0; pw < outputW; ++pw) {
+            int wstart = pw * strideW - paddingW;
+            int wend = std::min(wstart + sizeX, imgSizeW);
             wstart = std::max(wstart, 0);
-            dend = std::min(dend, static_cast<int>(inDepth));
-            hend = std::min(hend, static_cast<int>(inHeight));
-            wend = std::min(wend, static_cast<int>(inWidth));
 
-            CHECK(poolSize);
             tgtData[(pd * outputH + ph) * outputW + pw] = 0;  // clear
             for (int d = dstart; d < dend; ++d) {
               for (int h = hstart; h < hend; ++h) {
                 for (int w = wstart; w < wend; ++w) {
                   tgtData[(pd * outputH + ph) * outputW + pw] +=
-                      inData[(d * inHeight + h) * inWidth + w];
+                      inData[(d * imgSizeH + h) * imgSizeW + w];
                 }
               }
             }
+            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
+            CHECK(poolSize);
             tgtData[(pd * outputH + ph) * outputW + pw] /= poolSize;
           }
         }
       }
       // compute offset
-      inData += inDepth * inHeight * inWidth;
-      tgtData += outputD * outputH * outputW;
+      inData += inLength;
+      tgtData += outLength;
     }
   }
 }
@@ -2437,8 +2411,10 @@ void CpuMatrix::avgPool3DBackward(Matrix& input,
                                   real scaleTargets,
                                   real scaleOutput) {
   size_t num = input.getHeight();
-  size_t channels = input.getWidth() / outputD / outputH / outputW;
-  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth());
+  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
+  size_t outLength = outputH * outputW * outputD;
+  size_t channels = input.getWidth() / outLength;
+  CHECK(inLength * channels == getWidth());
   real* inData = input.getData();
   real* outData = getData();
 
@@ -2448,21 +2424,18 @@ void CpuMatrix::avgPool3DBackward(Matrix& input,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t pd = 0; pd < outputD; ++pd) {
-        for (size_t ph = 0; ph < outputH; ++ph) {
-          for (size_t pw = 0; pw < outputW; ++pw) {
         int dstart = pd * strideD - paddingD;
-            int hstart = ph * strideH - paddingH;
-            int wstart = pw * strideW - paddingW;
-            int dend = std::min(dstart + sizeZ, imgSizeD + paddingD);
-            int hend = std::min(hstart + sizeY, imgSizeH + paddingH);
-            int wend = std::min(wstart + sizeX, imgSizeW + paddingW);
-            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
+        int dend = std::min(dstart + sizeZ, imgSizeD);
         dstart = std::max(dstart, 0);
+        for (size_t ph = 0; ph < outputH; ++ph) {
+          int hstart = ph * strideH - paddingH;
+          int hend = std::min(hstart + sizeY, imgSizeH);
           hstart = std::max(hstart, 0);
+          for (size_t pw = 0; pw < outputW; ++pw) {
+            int wstart = pw * strideW - paddingW;
+            int wend = std::min(wstart + sizeX, imgSizeW);
             wstart = std::max(wstart, 0);
-            dend = std::min(dend, static_cast<int>(imgSizeD));
-            hend = std::min(hend, static_cast<int>(imgSizeH));
-            wend = std::min(wend, static_cast<int>(imgSizeW));
+            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
             CHECK(poolSize);
             for (int d = dstart; d < dend; ++d) {
               for (int h = hstart; h < hend; ++h) {
@@ -2476,8 +2449,8 @@ void CpuMatrix::avgPool3DBackward(Matrix& input,
         }
       }
       // offset
-      outData += imgSizeD * imgSizeH * imgSizeW;
-      inData += outputD * outputH * outputW;
+      outData += inLength;
+      inData += outLength;
     }
   }
 }

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/math/NEONFunctions.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
+
+namespace paddle {
+namespace neon {
+
+void relu(const float* a, float* b, int len);
+
+}  // namespace neon
+}  // namespace paddle

paddle/math/tests/test_matrixCompare.cpp

@@ -825,9 +825,8 @@ void testMaxPoolFwdBwd(int numSamples,
                        int strideW,
                        int padH,
                        int padW) {
-  int outH = 0, outW = 0;
-  outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1;
-  outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1;
+  int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true);
+  int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true);
 
   int inWidth = imgSizeH * imgSizeW * channels;
   MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
@@ -927,9 +926,8 @@ void testAvgPoolFwdBwd(int numSamples,
                        int strideW,
                        int padH,
                        int padW) {
-  int outH = 0, outW = 0;
-  outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1;
-  outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1;
+  int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true);
+  int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true);
 
   int inWidth = imgSizeH * imgSizeW * channels;
   MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);

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/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/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.
@@ -44,12 +44,13 @@ class IdentityOp : public NetOp {
       : NetOp(type, inputs, outputs, attrs) {
     PADDLE_ENFORCE_NE(Input("X"), framework::kEmptyVarName,
                       "Input(X) of IdentityOp should not be null.");
-    PADDLE_ENFORCE_NE(Output("Out"), framework::kEmptyVarName,
-                      "Output(Out) 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/minus_op.cc

@@ -71,7 +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);

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/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/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})

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/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)
 
 
@@ -179,8 +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
@@ -192,21 +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()]
@@ -241,8 +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_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_identity_op.py

@@ -7,13 +7,13 @@ class TestIdentityOp(OpTest):
     def setUp(self):
         self.op_type = "identity"
         self.inputs = {'X': np.random.random((10, 10)).astype("float32")}
-        self.outputs = {'Out': self.inputs['X']}
+        self.outputs = {'Y': self.inputs['X']}
 
     def test_check_output(self):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Out')
+        self.check_grad(['X'], 'Y')
 
 
 if __name__ == "__main__":

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()