Merge conflicts

CMakeLists.txt

@@ -13,7 +13,6 @@
 # limitations under the License
 
 cmake_minimum_required(VERSION 3.0)
-
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 set(PROJ_ROOT ${CMAKE_CURRENT_SOURCE_DIR})
 set(PROJ_BINARY_ROOT ${CMAKE_CURRENT_BINARY_DIR})

paddle/framework/CMakeLists.txt

@@ -5,6 +5,7 @@ nv_test(dim_test SRCS dim_test.cu DEPS ddim)
 
 cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory)
 cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
+cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
 
 cc_test(variable_test SRCS variable_test.cc)
 cc_test(scope_test SRCS scope_test.cc)

paddle/framework/attr_checker.h

@@ -4,6 +4,7 @@
 #include <functional>
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
 #include <vector>
 #include "paddle/platform/enforce.h"
 
@@ -41,6 +42,35 @@ class DefaultValueSetter {
   T default_value_;
 };
 
+template <typename T>
+class EnumInContainer {
+ public:
+  explicit EnumInContainer(const std::unordered_set<T>& c) : container_(c) {}
+  void operator()(T& val) const {
+    PADDLE_ENFORCE(container_.find(val) != container_.end(),
+                   "Value %s is not in enum container %s", val,
+                   ContainerDebugString());
+  }
+
+ private:
+  std::string ContainerDebugString() const {
+    std::ostringstream sout;
+    sout << "[";
+    size_t cnt = 0;
+    for (auto& v : container_) {
+      sout << v;
+      ++cnt;
+      if (cnt != container_.size()) {
+        sout << " ,";
+      }
+    }
+    sout << "]";
+    return sout.str();
+  }
+
+  std::unordered_set<T> container_;
+};
+
 // check whether a certain attribute fit its limits
 // an attribute can have more than one limits
 template <typename T>
@@ -50,6 +80,11 @@ class TypedAttrChecker {
  public:
   TypedAttrChecker(const std::string& attr_name) : attr_name_(attr_name) {}
 
+  TypedAttrChecker& InEnum(const std::unordered_set<T>& range) {
+    value_checkers_.push_back(EnumInContainer<T>(range));
+    return *this;
+  }
+
   TypedAttrChecker& LargerThan(const T& lower_bound) {
     value_checkers_.push_back(LargerThanChecker<T>(lower_bound));
     return *this;

paddle/framework/ddim.h

@@ -119,17 +119,6 @@ int arity(const DDim& ddim);
 
 std::ostream& operator<<(std::ostream&, const DDim&);
 
-template <int NDIMS>
-Eigen::DSizes<Eigen::DenseIndex, NDIMS> ToEigenDSizes(const DDim& dims) {
-  int rank = arity(dims);
-  PADDLE_ENFORCE(rank == NDIMS, "DDim and NDIMS must be same");
-  Eigen::DSizes<Eigen::DenseIndex, NDIMS> dsizes;
-  for (int d = 0; d < rank; d++) {
-    dsizes[d] = dims[d];
-  }
-  return dsizes;
-}
-
 }  // namespace framework
 }  // namespace paddle
 

paddle/framework/eigen.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "paddle/framework/tensor.h"
+#include "unsupported/Eigen/CXX11/Tensor"
+
+namespace paddle {
+namespace framework {
+
+// EigenDim converts paddle::platform::DDim into Eigen::DSizes.
+template <int D>
+struct EigenDim {
+  using Type = Eigen::DSizes<Eigen::DenseIndex, D>;
+
+  static Type From(const DDim& dims) {
+    PADDLE_ENFORCE(arity(dims) == D, "D must match arity(DDim)");
+    Type ret;
+    for (int d = 0; d < arity(dims); d++) {
+      ret[d] = dims[d];
+    }
+    return ret;
+  }
+};
+
+// Interpret paddle::platform::Tensor as EigenTensor and EigenConstTensor.
+template <typename T, size_t D, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+struct EigenTensor {
+  // TODO(qijun) Now, default type in unaligned, and we will make a benchmark on
+  // the speed of aligned and unaligned version in future.
+  using Type = Eigen::TensorMap<Eigen::Tensor<T, D, MajorType, IndexType>>;
+
+  using ConstType =
+      Eigen::TensorMap<Eigen::Tensor<const T, D, MajorType, IndexType>>;
+
+  static Type From(Tensor& tensor, DDim dims) {
+    return Type(tensor.data<T>(), EigenDim<D>::From(dims));
+  }
+
+  static Type From(Tensor& tensor) { return From(tensor, tensor.dims_); }
+
+  static ConstType From(const Tensor& tensor, DDim dims) {
+    return ConstType(tensor.data<T>(), EigenDim<D>::From(dims));
+  }
+
+  static ConstType From(const Tensor& tensor) {
+    return From(tensor, tensor.dims_);
+  }
+};
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+struct EigenVector : public EigenTensor<T, 1, MajorType, IndexType> {
+  // Flatten is to reshape a Tensor into a one dimension EigenVector
+  static typename EigenTensor<T, 1>::Type Flatten(Tensor& tensor) {
+    return EigenTensor<T, 1>::From(
+        tensor, make_ddim({static_cast<int>(product(tensor.dims_))}));
+  }
+
+  static typename EigenTensor<T, 1>::ConstType Flatten(const Tensor& tensor) {
+    return EigenTensor<T, 1>::From(
+        tensor, make_ddim({static_cast<int>(product(tensor.dims_))}));
+  }
+};
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = EigenTensor<T, 2, MajorType, IndexType>;
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/eigen_test.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/eigen.h"
+#include <gtest/gtest.h>
+
+namespace paddle {
+namespace framework {
+
+TEST(EigenDim, From) {
+  EigenDim<3>::Type ed = EigenDim<3>::From(make_ddim({1, 2, 3}));
+  ASSERT_EQ(1, ed[0]);
+  ASSERT_EQ(2, ed[1]);
+  ASSERT_EQ(3, ed[2]);
+}
+
+TEST(Eigen, Tensor) {
+  Tensor t;
+  float* p = t.mutable_data<float>(make_ddim({1, 2, 3}), platform::CPUPlace());
+  for (int i = 0; i < 1 * 2 * 3; i++) {
+    p[i] = static_cast<float>(i);
+  }
+
+  EigenTensor<float, 3>::Type et = EigenTensor<float, 3>::From(t);
+
+  ASSERT_EQ(1, et.dimension(0));
+  ASSERT_EQ(2, et.dimension(1));
+  ASSERT_EQ(3, et.dimension(2));
+
+  for (int i = 0; i < 1; i++) {
+    for (int j = 0; j < 2; j++) {
+      for (int k = 0; k < 3; k++) {
+        ASSERT_NEAR((i * 2 + j) * 3 + k, et(i, j, k), 1e-6f);
+      }
+    }
+  }
+}
+
+TEST(Eigen, VectorFrom) {
+  Tensor t;
+  float* p = t.mutable_data<float>(make_ddim({6}), platform::CPUPlace());
+  for (int i = 0; i < 6; i++) {
+    p[i] = static_cast<float>(i);
+  }
+
+  EigenVector<float>::Type ev = EigenVector<float>::From(t);
+
+  ASSERT_EQ(6, ev.dimension(0));
+
+  for (int i = 0; i < 6; i++) {
+    ASSERT_NEAR(i, ev(i), 1e-6f);
+  }
+}
+
+TEST(Eigen, VectorFlatten) {
+  Tensor t;
+  float* p = t.mutable_data<float>(make_ddim({1, 2, 3}), platform::CPUPlace());
+  for (int i = 0; i < 1 * 2 * 3; i++) {
+    p[i] = static_cast<float>(i);
+  }
+
+  EigenVector<float>::Type ev = EigenVector<float>::Flatten(t);
+
+  ASSERT_EQ(1 * 2 * 3, ev.dimension(0));
+
+  for (int i = 0; i < 1 * 2 * 3; i++) {
+    ASSERT_NEAR(i, ev(i), 1e-6f);
+  }
+}
+
+TEST(Eigen, Matrix) {
+  Tensor t;
+  float* p = t.mutable_data<float>(make_ddim({2, 3}), platform::CPUPlace());
+  for (int i = 0; i < 2 * 3; i++) {
+    p[i] = static_cast<float>(i);
+  }
+
+  EigenMatrix<float>::Type em = EigenMatrix<float>::From(t);
+
+  ASSERT_EQ(2, em.dimension(0));
+  ASSERT_EQ(3, em.dimension(1));
+
+  for (int i = 0; i < 2; i++) {
+    for (int j = 0; j < 3; j++) {
+      ASSERT_NEAR(i * 3 + j, em(i, j), 1e-6f);
+    }
+  }
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/net.cc

@@ -19,7 +19,10 @@
 namespace paddle {
 namespace framework {
 
-void PlainNet::CompleteAddOp() {
+void PlainNet::CompleteAddOp(bool calc) {
+  add_op_done_ = true;
+  if (!calc) return;
+
   std::unordered_set<std::string> input_set;
   std::unordered_set<std::string> output_set;
   std::unordered_set<std::string> temp_output;
@@ -52,7 +55,6 @@ void PlainNet::CompleteAddOp() {
   }
 
   attrs_["temporary_index"] = tmp_index;
-  add_op_done_ = true;
 }
 
 std::string PlainNet::DebugString() const {

paddle/framework/net.h

@@ -16,7 +16,6 @@ limitations under the License. */
 
 #include <paddle/framework/op_desc.pb.h>
 #include <paddle/framework/operator.h>
-#include "paddle/framework/net_proto.pb.h"
 #include "paddle/framework/op_proto.pb.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/scope.h"
@@ -41,7 +40,7 @@ namespace framework {
 class Net : public OperatorBase {
  public:
   virtual void AddOp(const OperatorPtr& op) = 0;
-  virtual void CompleteAddOp() = 0;
+  virtual void CompleteAddOp(bool calc) = 0;
 };
 
 using NetPtr = std::shared_ptr<Net>;
@@ -86,7 +85,7 @@ class PlainNet : public Net {
     ops_.push_back(op);
   }
 
-  void CompleteAddOp() override;
+  void CompleteAddOp(bool calculate = true) override;
 
   std::string DebugString() const override;
 

paddle/framework/tensor.h

@@ -35,6 +35,15 @@ struct CastToPyBufferImpl;
 namespace framework {
 
 class Tensor {
+  template <bool less, size_t i, typename... args>
+  friend struct paddle::pybind::details::CastToPyBufferImpl;
+
+  template <typename T, size_t D, int MajorType, typename IndexType>
+  friend struct EigenTensor;
+
+  template <typename T, int MajorType, typename IndexType>
+  friend struct EigenVector;
+
  public:
   Tensor() : offset_(0) {}
 
@@ -46,7 +55,7 @@ class Tensor {
   }
 
   template <typename T>
-  T* raw_data() const {
+  T* data() {
     CheckDims<T>();
     return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
                                 offset_);
@@ -86,66 +95,6 @@ class Tensor {
                                 offset_);
   }
 
-  template <typename T, size_t NDIMS>
-  typename TTypes<T, NDIMS>::Tensor shaped(DDim new_dims) {
-    Eigen::array<Eigen::DenseIndex, NDIMS> dims =
-        paddle::framework::ToEigenDSizes<NDIMS>(new_dims);
-    return typename TTypes<T, NDIMS>::Tensor(raw_data<T>(), dims);
-  }
-
-  template <typename T, size_t NDIMS>
-  typename TTypes<T, NDIMS>::Tensor tensor() {
-    return typename TTypes<T, NDIMS>::Tensor(
-        raw_data<T>(), paddle::framework::ToEigenDSizes<NDIMS>(dims_));
-  }
-
-  // flat to rank = 1
-  template <typename T>
-  typename TTypes<T>::Flat flat() {
-    return shaped<T, 1>(make_ddim({static_cast<int>(product(dims_))}));
-  }
-
-  // to TensorType Vec
-  template <typename T>
-  typename TTypes<T>::Vec vec() {
-    return tensor<T, 1>();
-  }
-
-  // to TensorType Matrix
-  template <typename T>
-  typename TTypes<T>::Matrix matrix() {
-    return tensor<T, 2>();
-  }
-
-  // const versions of all the methods above.
-  template <typename T, size_t NDIMS>
-  typename TTypes<T, NDIMS>::Tensor shaped(DDim new_dims) const {
-    Eigen::array<Eigen::DenseIndex, NDIMS> dims =
-        paddle::framework::ToEigenDSizes<NDIMS>(new_dims);
-    return typename TTypes<T, NDIMS>::Tensor(data<T>(), dims);
-  }
-
-  template <typename T, size_t NDIMS>
-  typename TTypes<T, NDIMS>::ConstantTensor tensor() const {
-    return typename TTypes<T, NDIMS>::Tensor(
-        data<T>(), paddle::framework::ToEigenDSizes<NDIMS>(dims_));
-  }
-
-  template <typename T>
-  typename TTypes<T>::ConstFlat flat() const {
-    return shaped<T, 1>(make_ddim({static_cast<int>(product(dims_))}));
-  }
-
-  template <typename T>
-  typename TTypes<T>::ConstVec vec() const {
-    return tensor<T, 1>();
-  }
-
-  template <typename T>
-  typename TTypes<T>::ConstMatrix matrix() const {
-    return tensor<T, 2>();
-  }
-
   template <typename T>
   void ShareDataFrom(const Tensor& src) {
     src.CheckDims<T>();
@@ -251,8 +200,6 @@ class Tensor {
   std::shared_ptr<Placeholder> holder_;  // holds the memory block if allocated.
   DDim dims_;
   size_t offset_;  // marks the begin of tensor data area.
-  template <bool less, size_t i, typename... args>
-  friend struct paddle::pybind::details::CastToPyBufferImpl;
 };
 
 }  // namespace framework

paddle/framework/tensor_types.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 "unsupported/Eigen/CXX11/Tensor"
-
-namespace paddle {
-namespace framework {
-
-// Helper to define Tensor types given that the scalar is of type T.
-template <typename T, int NDIMS = 1, typename IndexType = Eigen::DenseIndex>
-struct TTypes {
-  // Rank-<NDIMS> tensor of scalar type T.
-  typedef Eigen::TensorMap<Eigen::Tensor<T, NDIMS, Eigen::RowMajor, IndexType>,
-                           Eigen::Aligned>
-      Tensor;
-  typedef Eigen::TensorMap<
-      Eigen::Tensor<const T, NDIMS, Eigen::RowMajor, IndexType>, Eigen::Aligned>
-      ConstTensor;
-
-  // Scalar tensor (implemented as a rank-0 tensor) of scalar type T.
-  typedef Eigen::TensorMap<
-      Eigen::TensorFixedSize<T, Eigen::Sizes<>, Eigen::RowMajor, IndexType>,
-      Eigen::Aligned>
-      Scalar;
-  typedef Eigen::TensorMap<Eigen::TensorFixedSize<const T, Eigen::Sizes<>,
-                                                  Eigen::RowMajor, IndexType>,
-                           Eigen::Aligned>
-      ConstScalar;
-
-  // Rank-1 tensor (vector) of scalar type T.
-  typedef Eigen::TensorMap<Eigen::Tensor<T, 1, Eigen::RowMajor, IndexType>,
-                           Eigen::Aligned>
-      Flat;
-  typedef Eigen::TensorMap<
-      Eigen::Tensor<const T, 1, Eigen::RowMajor, IndexType>, Eigen::Aligned>
-      ConstFlat;
-  typedef Eigen::TensorMap<Eigen::Tensor<T, 1, Eigen::RowMajor, IndexType>,
-                           Eigen::Aligned>
-      Vec;
-  typedef Eigen::TensorMap<
-      Eigen::Tensor<const T, 1, Eigen::RowMajor, IndexType>, Eigen::Aligned>
-      ConstVec;
-
-  // Rank-2 tensor (matrix) of scalar type T.
-  typedef Eigen::TensorMap<Eigen::Tensor<T, 2, Eigen::RowMajor, IndexType>,
-                           Eigen::Aligned>
-      Matrix;
-  typedef Eigen::TensorMap<
-      Eigen::Tensor<const T, 2, Eigen::RowMajor, IndexType>, Eigen::Aligned>
-      ConstMatrix;
-};
-
-}  // namespace framework
-}  // namespace paddle

paddle/function/CMakeLists.txt

@@ -36,6 +36,7 @@ if(WITH_GPU)
     add_simple_unittest(MulOpTest)
     add_simple_unittest(CosSimOpTest)
     add_simple_unittest(RowConvOpTest)
+    add_simple_unittest(CropOpTest)
 endif()
 
 add_simple_unittest(ConvOpTest)

paddle/function/CropOp.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "CropOp.h"
+#include "paddle/function/TensorShape.h"
+#include "paddle/math/Vector.h"
+
+namespace paddle {
+
+template <>
+void Crop<DEVICE_TYPE_CPU>(real* outputs,
+                           const real* inputs,
+                           const TensorShape inShape,
+                           const TensorShape outShape,
+                           const FuncConfig& conf) {
+  std::vector<uint32_t> crop_corner =
+      conf.get<std::vector<uint32_t>>("crop_corner");
+  int cCrop = crop_corner[1];
+  int hCrop = crop_corner[2];
+  int wCrop = crop_corner[3];
+
+  int num = inShape[0];
+  int inC = inShape[1];
+  int inH = inShape[2];
+  int inW = inShape[3];
+
+  int outC = outShape[1];
+  int outH = outShape[2];
+  int outW = outShape[3];
+
+  for (int n = 0; n < num; n++) {
+    for (int c = 0; c < outC; c++) {
+      for (int h = 0; h < outH; h++) {
+        int outoff = ((n * outC + c) * outH + h) * outW;
+        int inoff = ((n * inC + c + cCrop) * inH + h + hCrop) * inW + wCrop;
+        memcpy(outputs + outoff, inputs + inoff, outW * sizeof(real));
+      }
+    }
+  }
+}
+
+template <>
+void CropGrad<DEVICE_TYPE_CPU>(const real* inGrad,
+                               real* outGrad,
+                               const TensorShape inShape,
+                               const TensorShape outShape,
+                               const FuncConfig& conf) {
+  std::vector<uint32_t> crop_corner =
+      conf.get<std::vector<uint32_t>>("crop_corner");
+  int cCrop = crop_corner[1];
+  int hCrop = crop_corner[2];
+  int wCrop = crop_corner[3];
+
+  int num = outShape[0];
+  int outC = outShape[1];
+  int outH = outShape[2];
+  int outW = outShape[3];
+
+  int inC = inShape[1];
+  int inH = inShape[2];
+  int inW = inShape[3];
+
+  for (int n = 0; n < num; n++) {
+    for (int c = 0; c < inC; c++) {
+      for (int h = 0; h < inH; h++) {
+        int outoff = ((n * outC + c + cCrop) * outH + h + hCrop) * outW + wCrop;
+        int inoff = ((n * inC + c) * inH + h) * inW;
+        CpuVector inG = CpuVector(inW, const_cast<real*>(inGrad + inoff));
+        CpuVector outG = CpuVector(inW, outGrad + outoff);
+        outG += inG;
+      }
+    }
+  }
+}
+
+/**
+ * \brief Crop input according to the specify corner and shape.
+ *        The input and output is a 4D tensor. In CropFunc, we only
+ *        crop the 2nd to 4th dimension.
+ *
+ * Argument in this Function:
+ * \param pad_    A struct object contains the cropping corner and shape.
+ * \param inputs  A 4D tensor, only one input.
+ * \param outputs A 4D tensor, the output value after cropping.
+ *
+ * For example,
+ * Input(2,2,2,3) = [
+ *                    [ [[1,2,3], [3,4,5]],
+ *                      [[2,3,5], [1,6,7]] ],
+ *                    [ [[4,3,1], [1,8,7]],
+ *                      [[3,8,9], [2,3,5]] ]
+ *                  ] # the input shape is (2,2,2,3)
+ *
+ * pad_: if corner = (0,1,1) and crop_shape = (2,1,2)
+ * Output(2,2,1,2) = [
+ *                    [ [[4,5]],
+ *                      [[6,7]] ],
+ *                    [ [[8,7]],
+ *                      [[3,5]] ]
+ *                  ] # the input shape is (2,2,2,3)
+ */
+template <DeviceType Device>
+class CropFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override { conf_ = config; }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(1UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+    CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO);
+
+    TensorShape inShape = inputs[0].shape();
+    TensorShape outShape = outputs[0].shape();
+
+    Crop<Device>(outputs[0].data<real>(),
+                 inputs[0].data<real>(),
+                 inShape,
+                 outShape,
+                 conf_);
+  }
+
+private:
+  FuncConfig conf_;
+};
+
+/**
+ * \brief The backward propagation of cropping Function.
+ *
+ * Argument in this Function:
+ * \param crop_    The same meaning as it in CropFunc.
+ * \param inputs  The gradient with respect to the output value of CropFunc.
+ * \param outputs The gradient with respect to the input value of CropFunc.
+ */
+
+template <DeviceType Device>
+class CropGradFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override { conf_ = config; }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(1UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+    CHECK_EQ(outputs[0].getArgType(), ADD_TO);
+
+    TensorShape outShape = outputs[0].shape();
+    TensorShape inShape = inputs[0].shape();
+
+    CropGrad<Device>(inputs[0].data<real>(),
+                     outputs[0].data<real>(),
+                     inShape,
+                     outShape,
+                     conf_);
+  }
+
+private:
+  FuncConfig conf_;
+};
+
+REGISTER_TYPED_FUNC(Crop, CPU, CropFunc);
+REGISTER_TYPED_FUNC(CropGrad, CPU, CropGradFunc);
+#ifndef PADDLE_ONLY_CPU
+REGISTER_TYPED_FUNC(Crop, GPU, CropFunc);
+REGISTER_TYPED_FUNC(CropGrad, GPU, CropGradFunc);
+#endif
+
+}  // namespace paddle

paddle/function/CropOp.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 "Function.h"
+
+namespace paddle {
+
+/**
+ * \brief  This funtion crops inputs according to the specify start point and
+ *shape.
+ *
+ * \param[out] outputs	save results.
+ * \param[in]  inputs	input data.
+ * \param[in]  inShape  the shape of input tensor.
+ * \param[in]  conf     the cropping config
+ */
+template <DeviceType Device>
+void Crop(real* outputs,
+          const real* inputs,
+          const TensorShape inShape,
+          const TensorShape outShape,
+          const FuncConfig& conf);
+
+/**
+ * \brief   Cropping operation backward.
+ *
+ * \param[out] inGrad	gradients of previous layer
+ * \param[in]  outGrad  output gradient
+ * \param[in]  inShape  the shape of input tensor.
+ * \param[in]  conf     the cropping config
+ */
+template <DeviceType Device>
+void CropGrad(const real* inGrad,
+              real* outGrad,
+              const TensorShape inShape,
+              const TensorShape outShape,
+              const FuncConfig& conf);
+}  // namespace paddle

paddle/function/CropOpGpu.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "hl_base.h"
+#include "CropOp.h"
+
+namespace paddle {
+
+__global__ void KeCrop(real* outputs, const real* inputs,
+                      int inC, int inH, int inW,
+                      int cropC, int cropH, int cropW,
+                      int outC, int outH, int outW, int nthreads) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int w = idx % outW;
+    const int h = (idx / outW) % outH;
+    const int c = (idx / outW / outH) % outC;
+    const int n = idx / outW / outH / outC;
+
+    const int off = ((n * inC + c + cropC) * inH + h + cropH) * inW + cropW + w;
+    outputs[idx] = inputs[off];
+  }
+}
+
+template <>
+void Crop<DEVICE_TYPE_GPU>(real* outputs,
+                          const real* inputs,
+						  const TensorShape inShape,
+						  const TensorShape outShape,
+                          const FuncConfig& conf) {
+  std::vector<uint32_t> crop_corner = conf.get<std::vector<uint32_t>>("crop_corner");
+  int cropC = crop_corner[1];
+  int cropH = crop_corner[2];
+  int cropW = crop_corner[3];
+
+  int num = inShape[0];
+  int inC = inShape[1];
+  int inH = inShape[2];
+  int inW = inShape[3];
+
+  int outC = outShape[1];
+  int outH = outShape[2];
+  int outW = outShape[3];
+
+  size_t nth = num * outC * outH * outW;
+  int blockSize = 1024;
+  int gridSize = (nth + blockSize - 1) / blockSize;
+
+  KeCrop<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>
+    (outputs, inputs, inC, inH, inW, cropC, cropH, cropW,
+     outC, outH, outW, nth);
+  CHECK_SYNC("Crop");
+}
+
+__global__ void KeCropDiff(const real* inGrad, real* outGrad,
+                          int inC, int inH, int inW,
+                          int cropC, int cropH, int cropW,
+                          int outC, int outH, int outW, int nthreads) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int w = idx % inW;
+    const int h = (idx / inW) % inH;
+    const int c = (idx / inW / inH) % inC;
+    const int n = idx / inW / inH / inC;
+
+    const int off = ((n * outC + c + cropC) * outH + h + cropH) * outW + cropW + w;
+
+    outGrad[off] += inGrad[idx];
+  }
+}
+
+template <>
+void CropGrad<DEVICE_TYPE_GPU>(const real* inGrad,
+                              real* outGrad,
+                              const TensorShape inShape,
+                              const TensorShape outShape,
+                              const FuncConfig& conf) {
+  std::vector<uint32_t> crop_corner = conf.get<std::vector<uint32_t>>("crop_corner");
+  int cropC = crop_corner[1];
+  int cropH = crop_corner[2];
+  int cropW = crop_corner[3];
+
+  int num = outShape[0];
+  int outC = outShape[1];
+  int outH = outShape[2];
+  int outW = outShape[3];
+
+  int inC = inShape[1];
+  int inH = inShape[2];
+  int inW = inShape[3];
+
+  size_t nth = num * inC * inH * inW;
+  int blockSize = 1024;
+  int gridSize = (nth + blockSize - 1) / blockSize;
+
+  KeCropDiff <<<gridSize, blockSize, 0, STREAM_DEFAULT>>>
+    (inGrad, outGrad, inC, inH, inW, cropC, cropH, cropW,
+     outC, outH, outW, nth);
+  CHECK_SYNC("CropGrad");
+}
+
+}  // namespace paddle

paddle/function/CropOpTest.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include "FunctionTest.h"
+
+namespace paddle {
+
+TEST(Crop, real) {
+  for (size_t numSamples : {5, 32}) {
+    for (size_t channels : {5, 5, 32}) {
+      for (size_t imgSizeH : {5, 33, 100}) {
+        for (size_t imgSizeW : {5, 32, 96}) {
+          VLOG(3) << " numSamples=" << numSamples << " channels=" << channels
+                  << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW;
+          for (bool test_grad : {false, true}) {
+            CpuGpuFuncCompare compare(
+                test_grad ? "CropGrad" : "Crop",
+                FuncConfig()
+                    .set<std::vector<uint32_t>>("crop_corner", {0, 1, 1, 1})
+                    .set<std::vector<uint32_t>>("crop_shape", {0, 2, 3, 3}));
+            TensorShape inDims{numSamples, channels, imgSizeH, imgSizeW};
+            TensorShape outDims{numSamples, 2, 3, 3};
+            compare.addInputs(
+                BufferArg(VALUE_TYPE_FLOAT, test_grad ? outDims : inDims));
+            compare.addOutputs(BufferArg(VALUE_TYPE_FLOAT,
+                                         test_grad ? inDims : outDims,
+                                         test_grad ? ADD_TO : ASSIGN_TO),
+                               test_grad ? ADD_TO : ASSIGN_TO);
+            compare.run();
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/CropLayer.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "CropLayer.h"
+#include "paddle/utils/Stat.h"
+namespace paddle {
+
+REGISTER_LAYER(crop, CropLayer);
+
+bool CropLayer::init(const LayerMap& layerMap,
+                     const ParameterMap& parameterMap) {
+  /* Initialize the basic parent class */
+  Layer::init(layerMap, parameterMap);
+  CHECK_LE(static_cast<int>(inputLayers_.size()), 2);
+  CHECK_GE(static_cast<int>(inputLayers_.size()), 1);
+  crop_axis_ = config_.axis();
+  for (int i = 0; i < config_.offset_size(); i++) {
+    crop_offsets_.push_back(config_.offset(i));
+  }
+
+  // 1. get input_0 shape
+  auto& input0_img_conf = config_.inputs(0).image_conf();
+  inDims_ = TensorShape({0,
+                         input0_img_conf.channels(),
+                         input0_img_conf.has_img_size_y()
+                             ? input0_img_conf.img_size_y()
+                             : input0_img_conf.img_size(),
+                         input0_img_conf.img_size()});
+  // 2. get target dims from config
+  if (config_.inputs_size() == 1) {
+    targetDims_ = TensorShape({config_.shape(0),
+                               config_.shape(1),
+                               config_.shape(2),
+                               config_.shape(3)});
+  } else {
+    // 2. get input_1 shape
+    auto& input1_img_conf = config_.inputs(1).image_conf();
+    targetDims_ = TensorShape({0,
+                               input1_img_conf.channels(),
+                               input1_img_conf.has_img_size_y()
+                                   ? input1_img_conf.img_size_y()
+                                   : input1_img_conf.img_size(),
+                               input1_img_conf.img_size()});
+  }
+
+  // 3. get final crop corner
+  int dimSize = 4;
+  crop_corner_ = {0, 0, 0, 0};
+  for (int i = 0; i < dimSize; i++) {
+    if (i >= crop_axis_) {
+      if (crop_offsets_.size() > 1) {
+        crop_corner_[i] = crop_offsets_[i - crop_axis_];
+      } else {
+        crop_corner_[i] = crop_offsets_[0];
+      }
+    }
+  }
+
+  outDims_ = TensorShape(4);
+
+  createFunction(
+      forward_, "Crop", FuncConfig().set("crop_corner", crop_corner_));
+  createFunction(
+      backward_, "CropGrad", FuncConfig().set("crop_corner", crop_corner_));
+
+  return true;
+}
+
+void CropLayer::setOutDims() {
+  MatrixPtr input = inputLayers_[1]->getOutputValue();
+  size_t batchSize = input->getHeight();
+  // get target dims from input_1
+  if (config_.inputs_size() == 2) {
+    targetDims_.setDim(0, batchSize);
+    int ch = config_.inputs(0).image_conf().channels();
+    if (ch != 0) targetDims_.setDim(1, ch);
+    int h = inputLayers_[1]->getOutput().getFrameHeight();
+    if (h != 0) targetDims_.setDim(2, h);
+    int w = inputLayers_[1]->getOutput().getFrameWidth();
+    if (w != 0) targetDims_.setDim(3, w);
+  }
+  // get final crop shape from target dims and crop axis
+  std::vector<uint32_t> crop_shape;
+  int dimSize = 4;
+  for (int i = 0; i < dimSize; i++) {
+    if (i >= crop_axis_) {
+      crop_shape.push_back(targetDims_[i]);
+    } else {
+      crop_shape.push_back(inDims_[i]);
+    }
+  }
+
+  outDims_.reshape(
+      {crop_shape[0], crop_shape[1], crop_shape[2], crop_shape[3]});
+  output_.setFrameHeight(crop_shape[2]);
+  output_.setFrameWidth(crop_shape[3]);
+}
+
+void CropLayer::setInDims() {
+  MatrixPtr input = inputLayers_[0]->getOutputValue();
+  size_t batchSize = input->getHeight();
+  inDims_.setDim(0, batchSize);
+  int h = inputLayers_[0]->getOutput().getFrameHeight();
+  if (h != 0) inDims_.setDim(2, h);
+  int w = inputLayers_[0]->getOutput().getFrameWidth();
+  if (w != 0) inDims_.setDim(3, w);
+}
+
+void CropLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  setInDims();
+  setOutDims();
+  int size = outDims_[1] * outDims_[2] * outDims_[3];
+  resetOutput(outDims_[0], size);
+  MatrixPtr outV = getOutputValue();
+  REGISTER_TIMER_INFO("CropForward", getName().c_str());
+
+  BufferArgs inputs;
+  BufferArgs outputs;
+  inputs.addArg(*getInputValue(0), inDims_);
+  outputs.addArg(*getOutputValue(), outDims_, ASSIGN_TO);
+  forward_[0]->calc(inputs, outputs);
+}
+
+void CropLayer::backward(const UpdateCallback& callback) {
+  (void)callback;
+  REGISTER_TIMER_INFO("CropBackward", getName().c_str());
+
+  BufferArgs inputs;
+  BufferArgs outputs;
+  inputs.addArg(*getOutputGrad(), outDims_);
+  outputs.addArg(*getInputGrad(0), inDims_, ADD_TO);
+  backward_[0]->calc(inputs, outputs);
+}
+}  // namespace paddle

paddle/gserver/layers/CropLayer.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 "Layer.h"
+
+namespace paddle {
+
+/**
+ * \brief  This layer crop input according to the specify conf.
+ *         input_0: input to be cropped
+ *         input_1: optional reference input
+ *         axis: start dimension to be croped
+ *         offset: offset of cropping  in each dimension
+ *         shape: if reference input layer was not setted,
+ *                  crop input as this shape conf
+ */
+class CropLayer : public Layer {
+public:
+  explicit CropLayer(const LayerConfig& config) : Layer(config) {}
+
+  ~CropLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+
+protected:
+  void setOutDims();
+  void setInDims();
+
+  int32_t crop_axis_;
+  std::vector<uint32_t> crop_offsets_;
+  std::vector<uint32_t> crop_corner_;
+  TensorShape inDims_;
+  TensorShape targetDims_;
+  TensorShape outDims_;
+};
+}  // namespace paddle

paddle/gserver/tests/CMakeLists.txt

@@ -56,7 +56,7 @@ add_test(NAME test_DetectionOutput
 add_unittest_without_exec(test_ConvUnify
     test_ConvUnify.cpp
     LayerGradUtil.cpp)
-    
+
 add_test(NAME test_ConvUnify
     COMMAND test_ConvUnify)
 ################# test_BatchNorm #######################

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1802,6 +1802,34 @@ TEST(Layer, RowConvLayer) {
   }
 }
 
+TEST(Layer, CropLayer) {
+  TestConfig config;
+  // config input_0
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 0});
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ImageConfig* img = input->mutable_image_conf();
+  img->set_channels(4);
+  img->set_img_size(16);
+  config.layerConfig.set_axis(2);
+  config.layerConfig.add_offset(0);
+  config.layerConfig.add_offset(0);
+
+  // config input_1
+  config.inputDefs.push_back({INPUT_DATA, "layer_1", 128, 0});
+  input = config.layerConfig.add_inputs();
+  img = input->mutable_image_conf();
+  img->set_channels(2);
+  img->set_img_size(8);
+
+  // config crop layer
+  config.layerConfig.set_type("crop");
+  config.layerConfig.set_name("cropLayer");
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "crop", 100, false, useGpu, false);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

paddle/operators/CMakeLists.txt

@@ -27,7 +27,8 @@ function(op_library TARGET)
     endif()
 
     list(LENGTH cu_srcs cu_srcs_len)
-    if (${cu_srcs_len} EQUAL 0)
+    list(LENGTH op_library_DEPS dep_len)
+    if (${cu_srcs_len} EQUAL 0 AND ${dep_len} EQUAL 0)
         message(WARNING "The op library ${TARGET} not support GPU!")
     endif()
 
@@ -47,3 +48,6 @@ op_library(mul_op SRCS mul_op.cc mul_op.cu)
 op_library(rowwise_add_op SRCS rowwise_add_op.cu rowwise_add_op.cc)
 op_library(sigmoid_op SRCS sigmoid_op.cu sigmoid_op.cc)
 op_library(softmax_op SRCS softmax_op.cc softmax_op.cu)
+
+op_library(fc_op SRCS fc_op.cc DEPS mul_op rowwise_add_op sigmoid_op
+        softmax_op net)

paddle/operators/add_op.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 #include "glog/logging.h"
+#include "paddle/framework/eigen.h"
 #include "paddle/framework/operator.h"
 
 namespace paddle {
@@ -29,8 +30,10 @@ public:
 
     output->mutable_data<T>(context.GetPlace());
 
-    output->flat<T>().device(*(context.GetEigenDevice<Place>())) =
-        input0.flat<T>() + input1.flat<T>();
+    framework::EigenVector<T>::Flatten(*output).device(
+        *(context.GetEigenDevice<Place>())) =
+        framework::EigenVector<T>::Flatten(input0) +
+        framework::EigenVector<T>::Flatten(input1);
   }
 };
 

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/net.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/framework/operator.h"
+
+namespace paddle {
+namespace operators {
+
+class FullyConnectedOp : public framework::PlainNet {
+public:
+  void Init() override {
+    AddOp(framework::OpRegistry::CreateOp("mul",
+                                          {
+                                              Input("X"), Input("W"),
+                                          },
+                                          {Output("before_act")},
+                                          {}));
+    auto b = Input("b");
+    if (b != framework::OperatorBase::EMPTY_VAR_NAME()) {
+      AddOp(framework::OpRegistry::CreateOp("rowwise_add",
+                                            {Output("before_act"), Input("b")},
+                                            {Output("before_act")},
+                                            {}));
+    }
+
+    auto activation = GetAttr<std::string>("activation");
+    AddOp(framework::OpRegistry::CreateOp(
+        activation, {Output("before_act")}, {Output("Y")}, {}));
+    CompleteAddOp(false);
+  }
+};
+
+class FullyConnectedOpMaker : public framework::OpProtoAndCheckerMaker {
+public:
+  FullyConnectedOpMaker(framework::OpProto *proto,
+                        framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "the input of fc operator");
+    AddInput("W", "the weight of fc operator");
+    AddInput("b", "the bias of fc operator");
+
+    AddOutput("Y", "the output of fc operator");
+    AddOutput(
+        "before_act", "the before activation output of fc operator", true);
+    AddAttr<std::string>("activation", "The activation key for fc layer")
+        .SetDefault("sigmoid")
+        .InEnum({"sigmoid", "softmax"});
+
+    //! TODO(yuyang18): Complete comment;
+    AddComment("FullyConnected Operator");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+USE_OP(mul);
+USE_OP(rowwise_add);
+USE_OP(sigmoid);
+USE_OP(softmax);
+
+REGISTER_OP(fc,
+            paddle::operators::FullyConnectedOp,
+            paddle::operators::FullyConnectedOpMaker);

paddle/pybind/CMakeLists.txt

 cc_library(paddle_pybind SHARED SRCS pybind.cc DEPS pybind python
-        add_op mul_op rowwise_add_op sigmoid_op softmax_op)
+        add_op fc_op)

paddle/pybind/pybind.cc

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include <Python.h>
 #include <paddle/framework/op_registry.h>
+#include <paddle/framework/operator.h>
 #include <paddle/framework/scope.h>
 #include <paddle/pybind/tensor_bind.h>
 #include <pybind11/numpy.h>
@@ -26,10 +27,7 @@ namespace py = pybind11;
 namespace pd = paddle::framework;
 
 USE_OP(add_two);
-USE_OP(softmax);
-USE_OP(mul);
-USE_OP(rowwise_add);
-USE_OP(sigmoid);
+USE_OP_WITHOUT_KERNEL(fc);
 
 PYBIND11_PLUGIN(core) {
   py::module m("core", "C++ core of Paddle Paddle");
@@ -53,7 +51,9 @@ PYBIND11_PLUGIN(core) {
              self.mutable_data<int>(paddle::platform::CPUPlace());
            })
       .def("set", paddle::pybind::PyTensorSetFromArray<float>)
-      .def("set", paddle::pybind::PyTensorSetFromArray<int>);
+      .def("set", paddle::pybind::PyTensorSetFromArray<int>)
+      .def("shape",
+           [](pd::Tensor& self) { return pd::vectorize(self.dims()); });
 
   py::class_<pd::Variable>(m, "Variable", R"DOC(Variable Class.
 
@@ -83,15 +83,16 @@ All parameter, weight, gradient are variables in Paddle.
 
   //! @note: Be careful! PyBind will return std::string as an unicode, not
   //! Python str. If you want a str object, you should cast them in Python.
-  m.def("get_all_op_protos", []() -> std::vector<std::string> {
+  m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
     auto& protos = pd::OpRegistry::protos();
-    std::vector<std::string> ret_values;
+    std::vector<py::bytes> ret_values;
     for (auto it = protos.begin(); it != protos.end(); ++it) {
       PADDLE_ENFORCE(it->second.IsInitialized(),
                      "OpProto must all be initialized");
-      ret_values.emplace_back();
-      PADDLE_ENFORCE(it->second.SerializeToString(&ret_values.back()),
+      std::string str;
+      PADDLE_ENFORCE(it->second.SerializeToString(&str),
                      "Serialize OpProto Error. This could be a bug of Paddle.");
+      ret_values.push_back(py::bytes(str));
     }
     return ret_values;
   });
@@ -101,17 +102,26 @@ All parameter, weight, gradient are variables in Paddle.
       .def("empty", pd::OperatorBase::EMPTY_VAR_NAME)
       .def("temp", pd::OperatorBase::TMP_VAR_NAME);
 
+  py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
+      .def_static("cpu_context", []() -> paddle::platform::DeviceContext* {
+        return new paddle::platform::CPUDeviceContext();
+      });
+
   py::class_<pd::OperatorBase, pd::OperatorPtr>(m, "Operator")
       .def("__str__", &pd::OperatorBase::DebugString)
-      .def_static("create", [](const std::string& protobin) {
-        pd::OpDesc desc;
-        PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
-                       "Cannot parse user input to OpDesc");
-        PADDLE_ENFORCE(desc.IsInitialized(),
-                       "User OpDesc is not initialized, reason %s",
-                       desc.InitializationErrorString());
-        return pd::OpRegistry::CreateOp(desc);
-      });
+      .def_static("create",
+                  [](py::bytes protobin) {
+                    pd::OpDesc desc;
+                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                                   "Cannot parse user input to OpDesc");
+                    PADDLE_ENFORCE(desc.IsInitialized(),
+                                   "User OpDesc is not initialized, reason %s",
+                                   desc.InitializationErrorString());
+                    return pd::OpRegistry::CreateOp(desc);
+                  })
+      .def("infer_shape", &pd::OperatorBase::InferShape)
+      .def("run", &pd::OperatorBase::Run)
+      .def("outputs", [](const pd::OperatorPtr& op) { return op->outputs_; });
 
   return m.ptr();
 }

proto/ModelConfig.proto

@@ -472,10 +472,16 @@ message LayerConfig {
   // blank label used in ctc loss
   optional uint32 blank = 52 [default = 0];
 
-  // stride parameter for seqlastins layer, AverageLayer, MaxLayer, which 
+  // stride parameter for seqlastins layer, AverageLayer, MaxLayer, which
   // controls the scope of pooling operation. can be set > 0.
   // leave empty or set to -1 to disable this stride pooling.
   optional int32 seq_pool_stride = 53 [default = -1];
+
+  // for crop layer
+  optional int32 axis = 54 [default = 2];
+  repeated uint32 offset = 55;
+  repeated uint32 shape = 56;
+
 }
 
 message EvaluatorConfig {

python/paddle/trainer/config_parser.py

@@ -1998,6 +1998,23 @@ class PadLayer(LayerBase):
         self.config.size = out_ch * out_h * out_w
 
 
+@config_layer('crop')
+class CropLayer(LayerBase):
+    def __init__(self, name, inputs, axis, offset, shape, **xargs):
+        super(CropLayer, self).__init__(name, 'crop', 0, inputs=inputs, **xargs)
+        self.config.axis = axis
+        self.config.offset.extend(offset)
+        self.config.shape.extend(shape)
+
+        # get channel, width and height from input_0 layer
+        input_layer = self.get_input_layer(0)
+        image_conf = self.config.inputs[0].image_conf
+        image_conf.img_size = input_layer.width
+        image_conf.img_size_y = input_layer.height
+        image_conf.channels = input_layer.size / (input_layer.width *
+                                                  input_layer.height)
+
+
 @config_layer('batch_norm')
 class BatchNormLayer(LayerBase):
     layer_type = 'batch_norm'

python/paddle/trainer_config_helpers/layers.py

@@ -127,6 +127,7 @@ __all__ = [
     'dropout_layer',
     'prelu_layer',
     'gated_unit_layer',
+    'crop_layer',
 ]
 
 
@@ -218,6 +219,7 @@ class LayerType(object):
     SMOOTH_L1 = 'smooth_l1'
 
     PRELU = 'prelu'
+    CROP_LAYER = 'crop'
 
     @staticmethod
     def is_layer_type(type_name):
@@ -5970,3 +5972,52 @@ def gated_unit_layer(input,
         name="%s_gated_act" % name,
         input=dotmul_operator(input_proj, gate),
         layer_attr=layer_attr)
+
+
+@wrap_name_default()
+@layer_support()
+def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None):
+    """
+    The crop layer crops images by offset and shape. User can set crop shape by
+    args 'shape' explicitly or by reference input layer.
+    
+    The example usage is:
+
+    .. code-block:: python
+    crop = crop_layer(input=[image_input, reference_input], axis=2, offset=[2, 3])
+
+    :param input: The input layer.If two inputs were setted,
+                    the second input will be regarded as reference input
+    :type input: LayerOutput or Sequence
+    :param offset: The crop offset
+    :type offset: Sequence
+    :param axis: start axis to be cropped. To image input layer:
+        - 0: batch size
+        - 1: channels
+        - 2: height
+        - 3: width
+    :type partial_sum: int
+    :param shape: The shape to be cropped. Default is None.
+    :type shape: Sequence | None
+    :param name: Name of this layer.
+    :type name: basestring
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+    if isinstance(input, LayerOutput):
+        input = [input]
+    else:
+        assert isinstance(input, collections.Sequence)
+    l = Layer(
+        inputs=[x.name for x in input],
+        axis=axis,
+        offset=offset,
+        shape=shape,
+        name=name,
+        type=LayerType.CROP_LAYER,
+        **ExtraLayerAttribute.to_kwargs(layer_attr))
+    return LayerOutput(
+        name=name,
+        layer_type=LayerType.CROP_LAYER,
+        parents=input,
+        size=l.config.size)

python/paddle/trainer_config_helpers/tests/configs/test_crop.py

+from paddle.trainer_config_helpers import *
+
+settings(batch_size=1000, learning_rate=1e-5)
+
+data = data_layer(name='data', size=2016, height=48, width=42)
+refernce_data = data_layer(name='data', size=768, height=16, width=16)
+
+conv = img_conv_layer(
+    input=data,
+    filter_size=3,
+    num_channels=1,
+    num_filters=16,
+    padding=1,
+    act=LinearActivation(),
+    bias_attr=True)
+
+pool = img_pool_layer(input=conv, pool_size=2, stride=2, pool_type=MaxPooling())
+
+crop = crop_layer(input=[pool, refernce_data], axis=2)
+
+outputs(pad)

python/paddle/v2/framework/create_op_creation_methods.py

@@ -217,6 +217,10 @@ def create_op_creation_method(op_proto):
         return core.Operator.create(opdesc.SerializeToString())
 
     __impl__.__doc__ = get_docstring_from_op_proto(op_proto)
+    __impl__.all_input_args = [var.name for var in op_proto.inputs]
+    __impl__.all_output_args = [var.name for var in op_proto.outputs]
+    __impl__.all_attr_args = [attr.name for attr in op_proto.attrs]
+
     return __impl__
 
 

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

 add_python_test(test_framework test_protobuf.py test_scope.py
     test_default_scope_funcs.py test_op_creation_methods.py
-    test_tensor.py)
+    test_tensor.py test_fc_op.py test_add_two_op.py)

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

+import paddle.v2.framework.core as core
+import unittest
+import numpy
+import paddle.v2.framework.create_op_creation_methods as creation
+
+
+class OpTestMeta(type):
+    def __new__(cls, name, bases, attrs):
+        obj = super(OpTestMeta, cls).__new__(cls, name, bases, attrs)
+
+        def test_all(self):
+            func = getattr(creation.op_creations, self.type, None)
+            self.assertIsNotNone(func)
+
+            scope = core.Scope(None)
+            kwargs = dict()
+
+            for in_name in func.all_input_args:
+                if hasattr(self, in_name):
+                    kwargs[in_name] = in_name
+                    var = scope.create_var(in_name).get_tensor()
+                    arr = getattr(self, in_name)
+                    var.set_dims(arr.shape)
+                    var.set(arr)
+                else:
+                    kwargs[in_name] = "@EMPTY@"
+
+            for out_name in func.all_output_args:
+                if hasattr(self, out_name):
+                    kwargs[out_name] = out_name
+                    scope.create_var(out_name).get_tensor()
+
+            for attr_name in func.all_attr_args:
+                if hasattr(self, attr_name):
+                    kwargs[attr_name] = getattr(self, attr_name)
+
+            op = func(**kwargs)
+
+            op.infer_shape(scope)
+
+            ctx = core.DeviceContext.cpu_context()
+            op.run(scope, ctx)
+
+            for out_name in func.all_output_args:
+                actual = numpy.array(scope.get_var(out_name).get_tensor())
+                expect = getattr(self, out_name)
+                numpy.testing.assert_almost_equal(actual, expect)
+
+        obj.test_all = test_all
+        return obj

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

+import unittest
+from op_test_util import OpTestMeta
+import numpy
+
+
+class TestAddOp(unittest.TestCase):
+    __metaclass__ = OpTestMeta
+
+    def setUp(self):
+        self.type = "add_two"
+        self.X = numpy.random.random((342, 345)).astype("float32")
+        self.Y = numpy.random.random((342, 345)).astype("float32")
+        self.Out = self.X + self.Y
+
+
+if __name__ == '__main__':
+    unittest.main()

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

+import paddle.v2.framework.core as core
+import unittest
+import numpy
+import paddle.v2.framework.create_op_creation_methods as creation
+
+
+class TestFc(unittest.TestCase):
+    def test_fc(self):
+        scope = core.Scope(None)
+        x = scope.create_var("X")
+        x_tensor = x.get_tensor()
+        x_tensor.set_dims([1000, 784])
+        x_tensor.alloc_float()
+
+        w = scope.create_var("W")
+        w_tensor = w.get_tensor()
+        w_tensor.set_dims([784, 100])
+        w_tensor.alloc_float()
+
+        w_tensor.set(numpy.random.random((784, 100)).astype("float32"))
+
+        # Set a real numpy array here.
+        # x_tensor.set(numpy.array([]))
+
+        op = creation.op_creations.fc(X="X", Y="Y", W="W")
+
+        for out in op.outputs():
+            if scope.get_var(out) is None:
+                scope.create_var(out).get_tensor()
+
+        tensor = scope.get_var("Y").get_tensor()
+        op.infer_shape(scope)
+        self.assertEqual([1000, 100], tensor.shape())
+
+        ctx = core.DeviceContext.cpu_context()
+
+        op.run(scope, ctx)
+
+        # After complete all ops, check Y is expect or not.
+
+
+if __name__ == '__main__':
+    unittest.main()