fix conflicts

Dockerfile

@@ -27,7 +27,7 @@ RUN apt-get update && \
     git python-pip python-dev openssh-server bison  \
     wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
     curl sed grep graphviz libjpeg-dev zlib1g-dev  \
-    python-numpy python-matplotlib gcc g++ \
+    python-numpy python-matplotlib gcc-4.8 g++-4.8 \
     automake locales clang-format-3.8 swig doxygen cmake  \
     liblapack-dev liblapacke-dev libboost-dev \
     clang-3.8 llvm-3.8 libclang-3.8-dev \

cmake/flags.cmake

@@ -9,6 +9,11 @@ function(CheckCompilerCXX11Flag)
         if(${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 4.8)
             message(FATAL_ERROR "Unsupported GCC version. GCC >= 4.8 required.")
         endif()
+        # TODO(qijun) gcc 4.9 or later versions raise SEGV due to the optimization problem.
+        # Use Debug mode instead for now.
+        if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 4.9 OR CMAKE_CXX_COMPILER_VERSION VERSION_EQUAL 4.9) 
+            set(CMAKE_BUILD_TYPE "Debug" CACHE STRING "" FORCE)
+        endif()
     elseif(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
         # cmake >= 3.0 compiler id "AppleClang" on Mac OS X, otherwise "Clang"
         # Apple Clang is a different compiler than upstream Clang which havs different version numbers.

doc/api/v2/config/layer.rst

@@ -104,6 +104,11 @@ cross_channel_norm
 ------------------
 ..  autoclass:: paddle.v2.layer.cross_channel_norm
     :noindex:
+
+row_l2_norm
+-----------
+..  autoclass:: paddle.v2.layer.row_l2_norm
+    :noindex:
     
 Recurrent Layers
 ================
@@ -320,6 +325,11 @@ scaling
 ..  autoclass:: paddle.v2.layer.scaling
     :noindex:
 
+clip
+----
+..  autoclass:: paddle.v2.layer.clip
+    :noindex:
+
 slope_intercept
 ---------------
 ..  autoclass:: paddle.v2.layer.slope_intercept

paddle/framework/detail/tensor-inl.h

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-
 #include "paddle/memory/memcpy.h"
 
 namespace paddle {
@@ -62,9 +61,11 @@ inline T* Tensor::mutable_data(platform::Place place) {
     if (platform::is_cpu_place(place)) {
       holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>(
           boost::get<platform::CPUPlace>(place), size));
+    } else if (platform::is_gpu_place(place)) {
+#ifdef PADDLE_ONLY_CPU
+      PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
     }
-#ifndef PADDLE_ONLY_CPU
+#else
-    else if (platform::is_gpu_place(place)) {
       holder_.reset(new PlaceholderImpl<T, platform::GPUPlace>(
           boost::get<platform::GPUPlace>(place), size));
     }

paddle/framework/operator.cc

@@ -20,16 +20,16 @@ namespace paddle {
 namespace framework {
 
 template <>
-Eigen::DefaultDevice* ExecutionContext::GetEigenDevice<
+Eigen::DefaultDevice& ExecutionContext::GetEigenDevice<
     platform::CPUPlace, Eigen::DefaultDevice>() const {
-  return device_context_.get_eigen_device<Eigen::DefaultDevice>();
+  return *device_context_.get_eigen_device<Eigen::DefaultDevice>();
 }
 
 #ifndef PADDLE_ONLY_CPU
 template <>
-Eigen::GpuDevice*
+Eigen::GpuDevice&
 ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
-  return device_context_.get_eigen_device<Eigen::GpuDevice>();
+  return *device_context_.get_eigen_device<Eigen::GpuDevice>();
 }
 #endif
 

paddle/framework/operator.h

@@ -253,7 +253,7 @@ class ExecutionContext : public OperatorContext {
   template <typename PlaceType,
             typename DeviceType =
                 typename EigenDeviceConverter<PlaceType>::EigenDeviceType>
-  DeviceType* GetEigenDevice() const;
+  DeviceType& GetEigenDevice() const;
 
   platform::Place GetPlace() const { return device_context_.GetPlace(); }
 

paddle/gserver/layers/ClipLayer.cpp

@@ -0,0 +1,79 @@
+/* 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 "Layer.h"
+
+namespace paddle {
+
+/**
+ * A layer for clipping the input value by the threshold.
+ * \f[
+ *   out[i] = \min\left(\max\left(in[i],p_{1}\right),p_{2}\right)
+ * \f]
+ */
+
+class ClipLayer : public Layer {
+protected:
+  double min_;
+  double max_;
+
+public:
+  explicit ClipLayer(const LayerConfig& config) : Layer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+};
+
+REGISTER_LAYER(clip, ClipLayer);
+
+bool ClipLayer::init(const LayerMap& layerMap,
+                     const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+
+  CHECK_EQ(inputLayers_.size(), 1U);
+  auto layerConf = config_.inputs(0).clip_conf();
+  min_ = layerConf.min();
+  max_ = layerConf.max();
+  CHECK_LT(min_, max_);
+  return true;
+}
+
+void ClipLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  MatrixPtr inV = getInputValue(0);
+  resetOutput(inV->getHeight(), inV->getWidth());
+  MatrixPtr outV = getOutputValue();
+  outV->copyFrom(*inV);
+  outV->clip(min_, max_);
+}
+
+void ClipLayer::backward(const UpdateCallback& callback) {
+  MatrixPtr inV = getInputValue(0);
+  MatrixPtr inG = getInputGrad(0);
+  if (inG) {
+    MatrixPtr outV = getOutputValue();
+    MatrixPtr outG = getOutputGrad();
+    MatrixPtr tmpMtx;
+    Matrix::resizeOrCreate(
+        tmpMtx, outG->getHeight(), outG->getWidth(), false, useGpu_);
+    tmpMtx->clipDerivative(*inV, min_, max_);
+    inG->addDotMul(*outG, *tmpMtx, 1, 1);
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/RowL2NormLayer.cpp

@@ -0,0 +1,98 @@
+/* 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 "Layer.h"
+
+namespace paddle {
+
+/**
+ * A layer for L2 normalization in each row,
+ * \f[
+ *   out[i] = \frac{in[i]}{\sqrt{\sum_{k=1}^N in[k]^{2}}}
+ * \f]
+ * where the size of \f$in\f$ is (batchSize x dataDim),
+ * and the size of \f$out\f$ is (batchSize x dataDim).
+ */
+
+class RowL2NormLayer : public Layer {
+protected:
+  MatrixPtr inSquare_;
+  MatrixPtr l2NormReciprocal_;
+  MatrixPtr dotSum_;
+
+public:
+  explicit RowL2NormLayer(const LayerConfig& config) : Layer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+};
+
+REGISTER_LAYER(row_l2_norm, RowL2NormLayer);
+
+bool RowL2NormLayer::init(const LayerMap& layerMap,
+                          const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+
+  CHECK_EQ(inputLayers_.size(), 1U);
+
+  return true;
+}
+
+void RowL2NormLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  MatrixPtr inV = getInputValue(0);
+
+  /* malloc memory for the output_ if necessary */
+  size_t batchSize = inV->getHeight();
+  size_t dataDim = getSize();
+  CHECK_EQ(dataDim, inV->getWidth());
+  resetOutput(batchSize, dataDim);
+  MatrixPtr outV = getOutputValue();
+
+  Matrix::resizeOrCreate(inSquare_, batchSize, dataDim, false, useGpu_);
+  inV->square2(*inSquare_);
+  Matrix::resizeOrCreate(l2NormReciprocal_, batchSize, 1, false, useGpu_);
+  inSquare_->rowSum(*l2NormReciprocal_);
+  l2NormReciprocal_->sqrt2(*l2NormReciprocal_);
+  l2NormReciprocal_->scalarDiv(*l2NormReciprocal_, 1.0);
+  outV->rowScale(0, *inV, *l2NormReciprocal_);
+}
+
+void RowL2NormLayer::backward(const UpdateCallback& callback) {
+  MatrixPtr inV = getInputValue(0);
+  MatrixPtr inG = getInputGrad(0);
+  MatrixPtr outV = getOutputValue();
+  MatrixPtr outG = getOutputGrad();
+  size_t batchSize = inV->getHeight();
+
+  // inG[ij] += outG[ij] / l2NormReciprocal
+  // inG[ij] += -inV[ij] * l2NormReciprocal * l2NormReciprocal * DotMul(outG[i],
+  // inV[i])
+  if (inG) {
+    Matrix::resizeOrCreate(dotSum_, batchSize, 1, false, useGpu_);
+    dotSum_->zeroMem();
+    dotSum_->rowDotMul(0, *outG, *outV);
+    dotSum_->dotMul(*dotSum_, *l2NormReciprocal_);
+    dotSum_->dotMul(*dotSum_, *l2NormReciprocal_);
+    inSquare_->rowScale(0, *inV, *dotSum_);
+    inG->sub(*inSquare_);
+    inG->addRowScale(0, *outG, *l2NormReciprocal_);
+  }
+}
+
+}  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1899,6 +1899,36 @@ TEST(Layer, CropLayer) {
   }
 }
 
+TEST(Layer, ClipLayer) {
+  const size_t batchSize = 128;
+  const size_t size = 512;
+  TestConfig config;
+  config.layerConfig.set_type("clip");
+  config.inputDefs.push_back({INPUT_DATA, "input", size, 0});
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ClipConfig* layerConf = input->mutable_clip_conf();
+  double p1 = std::rand() / (double)RAND_MAX;
+  double p2 = std::rand() / (double)RAND_MAX;
+  layerConf->set_min(std::min(p1, p2));
+  layerConf->set_max(std::max(p1, p2));
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "clip", batchSize, false, useGpu, false);
+  }
+}
+
+TEST(Layer, RowL2NormLayer) {
+  const size_t batchSize = 128;
+  const size_t size = 512;
+  TestConfig config;
+  config.layerConfig.set_type("row_l2_norm");
+  config.layerConfig.set_size(size);
+  config.inputDefs.push_back({INPUT_DATA, "input", size, 0});
+  config.layerConfig.add_inputs();
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "row_l2_norm", batchSize, false, useGpu, false);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

paddle/math/BaseMatrix.cu

@@ -442,6 +442,12 @@ DEFINE_MATRIX_UNARY_PARAMETER_OP(Clip, TWO_PARAMETER,
 template<class T>
 void BaseMatrixT<T>::clip(T p1, T p2) { applyUnary(unary::Clip<T>(p1, p2)); }
 
+DEFINE_MATRIX_BINARY_PARAMETER_OP(ClipDerivative, TWO_PARAMETER, a = b < p1 ? 0 : (b > p2 ? 0 : 1));
+template<class T>
+void BaseMatrixT<T>::clipDerivative(BaseMatrixT& b, T p1, T p2) {
+  applyBinary(binary::ClipDerivative<T>(p1, p2), b);
+}
+
 DEFINE_MATRIX_UNARY_PARAMETER_OP(BiggerThanScalar, ONE_PARAMETER,
                                  a = a > p ? 1.0f : 0.0f);
 template<class T>

paddle/math/BaseMatrix.h

@@ -488,6 +488,13 @@ public:
    */
   void clip(T p1, T p2);
 
+  /**
+   * this = b < low ? 0 : 1
+   *
+   * this = b > high ? 0 : 1
+   */
+  void clipDerivative(BaseMatrixT& b, T p1, T p2);
+
   /**
    * @code
    * a = a > p ? 1.0f : 0.0f

paddle/operators/add_op.cc

@@ -50,10 +50,6 @@ The equation is: Out = X + Y
 class AddOpGrad : public OperatorWithKernel {
 protected:
   void InferShape(const InferShapeContext &ctx) const override {}
-  std::string DebugString() const override {
-    LOG(INFO) << "AddOpGrad";
-    return "";
-  }
 };
 
 }  // namespace operators

paddle/operators/add_op.cu

@@ -1,3 +1,4 @@
+#define EIGEN_USE_GPU
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/add_op.h"
 

paddle/operators/add_op.h

@@ -32,7 +32,7 @@ public:
     auto Y = EigenVector<T>::Flatten(*input1);
     auto Z = EigenVector<T>::Flatten(*output);
 
-    auto place = *context.GetEigenDevice<Place>();
+    auto place = context.GetEigenDevice<Place>();
 
     Z.device(place) = X + Y;
   }

paddle/operators/cross_entropy_op.cu

@@ -1,3 +1,4 @@
+#define EIGEN_USE_GPU
 #include "paddle/operators/cross_entropy_op.h"
 
 REGISTER_OP_GPU_KERNEL(onehot_cross_entropy,

paddle/operators/mean_op.h

@@ -29,7 +29,7 @@ public:
 
     auto X = EigenVector<T>::Flatten(*input);
     auto y = EigenScalar<T>::From(*output);
-    auto place = *context.GetEigenDevice<Place>();
+    auto place = context.GetEigenDevice<Place>();
 
     y.device(place) = X.mean();
   }

paddle/operators/mul_op.cu

@@ -12,6 +12,7 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
+#define EIGEN_USE_GPU
 #include "paddle/operators/mul_op.h"
 
 REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<ops::GPUPlace, float>);

paddle/operators/mul_op.h

@@ -35,7 +35,7 @@ public:
     auto X = EigenMatrix<T>::From(*input0);
     auto Y = EigenMatrix<T>::From(*input1);
     auto Z = EigenMatrix<T>::From(*output);
-    auto place = *context.GetEigenDevice<Place>();
+    auto place = context.GetEigenDevice<Place>();
 
     Z.device(place) = X.contract(Y, dim_pair);
   }

paddle/operators/rowwise_add_op.cu

@@ -1,3 +1,4 @@
+#define EIGEN_USE_GPU
 #include "paddle/operators/rowwise_add_op.h"
 
 REGISTER_OP_GPU_KERNEL(rowwise_add,

paddle/operators/rowwise_add_op.h

@@ -33,7 +33,7 @@ public:
     const int rest_size = input.size() / bias_size;
     Eigen::DSizes<int, 1> one_d(input.size());
     Eigen::DSizes<int, 1> bcast(rest_size);
-    output.reshape(one_d).device(*(context.GetEigenDevice<Place>())) =
+    output.reshape(one_d).device(context.GetEigenDevice<Place>()) =
         input.reshape(one_d) + bias.broadcast(bcast).reshape(one_d);
   }
 };

paddle/operators/sgd_op.cu

@@ -1,3 +1,4 @@
+#define EIGEN_USE_GPU
 #include "paddle/operators/sgd_op.h"
 
 REGISTER_OP_GPU_KERNEL(sgd, ops::SGDOpKernel<ops::GPUPlace, float>);

paddle/operators/sgd_op.h

@@ -32,7 +32,7 @@ public:
     auto p = EigenVector<T>::Flatten(*param);
     auto g = EigenVector<T>::Flatten(*grad);
     auto o = EigenVector<T>::Flatten(*param_out);
-    auto place = *ctx.GetEigenDevice<Place>();
+    auto place = ctx.GetEigenDevice<Place>();
 
     o.device(place) = p - lr * g;
   }

paddle/operators/sigmoid_op.cu

@@ -1,3 +1,4 @@
+#define EIGEN_USE_GPU
 #include "paddle/operators/sigmoid_op.h"
 
 REGISTER_OP_GPU_KERNEL(sigmoid, ops::SigmoidKernel<ops::GPUPlace, float>);

paddle/operators/sigmoid_op.h

@@ -29,7 +29,7 @@ public:
 
     auto X = EigenVector<T>::Flatten(*input);
     auto Y = EigenVector<T>::Flatten(*output);
-    auto place = *context.GetEigenDevice<Place>();
+    auto place = context.GetEigenDevice<Place>();
 
     Y.device(place) = 1.0 / (1.0 + (-1.0 * X).exp());
   }

paddle/operators/softmax_op.cu

@@ -1,3 +1,4 @@
+#define EIGEN_USE_GPU
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/softmax_op.h"