Merge pull request #300 from QiJune/feature/sppnet

add SpatialPyramidPoolLayer c++ support

doc/ui/api/trainer_config_helpers/layers.rst

@@ -46,6 +46,12 @@ conv_operator
     :members: conv_operator
     :noindex:
 
+conv_projection
+-------------
+..  automodule:: paddle.trainer_config_helpers.layers
+    :members: conv_projection
+    :noindex:
+
 conv_shift_layer
 ------------------
 ..  automodule:: paddle.trainer_config_helpers.layers
@@ -73,6 +79,12 @@ img_pool_layer
     :members: img_pool_layer
     :noindex:   
 
+spp_layer
+--------------
+..  automodule:: paddle.trainer_config_helpers.layers
+    :members: spp_layer
+    :noindex:
+
 maxout_layer
 ------------
 ..  automodule:: paddle.trainer_config_helpers.layers

paddle/cuda/include/hl_cnn.h

@@ -91,6 +91,7 @@ extern void hl_expand_feature2col(
  * @param[in]   paddingH    padding height.
  * @param[in]   paddingW    padding width.
  * @param[out]  tgtData     output data.
+ * @param[in]   tgtStride   stride between output data samples.
  *
  */
 extern void hl_maxpool_forward(
@@ -100,7 +101,8 @@ extern void hl_maxpool_forward(
     const int pooledH, const int pooledW,
     const int sizeX, const int sizeY,
     const int strideH, const int strideW,
-    const int paddingH, const int paddingW, real* tgtData);
+    const int paddingH, const int paddingW,
+    real* tgtData, const int tgtStride);
 
 /**
  * @brief   Maximum pool backward.
@@ -123,6 +125,7 @@ extern void hl_maxpool_forward(
  * @param[in]   paddingH    padding height.
  * @param[in]   paddingW    padding width.
  * @param[out]  targetGrad  output grad.
+ * @param[in]   outStride   stride between output data samples. 
  *
  */
 extern void hl_maxpool_backward(
@@ -135,7 +138,7 @@ extern void hl_maxpool_backward(
     const int strideH, const int strideW,
     const int paddingH, const int paddingW,
     real scaleA, real scaleB,
-    real* targetGrad);
+    real* targetGrad, const int outStride);
 
 /**
  * @brief   Averge pool forward.
@@ -154,6 +157,7 @@ extern void hl_maxpool_backward(
  * @param[in]   paddingH    padding height.
  * @param[in]   paddingW    padding width.
  * @param[out]  tgtData     output data.
+ * @param[in]   tgtStride   stride between output data samples.
  *
  */
 extern void hl_avgpool_forward(
@@ -163,7 +167,8 @@ extern void hl_avgpool_forward(
     const int pooledH, const int pooledW,
     const int sizeX, const int sizeY,
     const int strideH, const int strideW,
-    const int paddingH, const int paddingW, real* tgtData);
+    const int paddingH, const int paddingW,
+    real* tgtData, const int tgtStride);
 
 /**
  * @brief   Maximum pool backward.
@@ -184,6 +189,7 @@ extern void hl_avgpool_forward(
  * @param[in]   scaleA      scale.
  * @param[in]   scaleB      scale.
  * @param[out]  backGrad    output grad.
+ * @param[in]   outStride   stride between output data samples. 
  *
  */
 extern void hl_avgpool_backward(
@@ -195,7 +201,7 @@ extern void hl_avgpool_backward(
     const int strideH, const int strideW,
     int paddingH, int paddingW,
     real scaleA, real scaleB,
-    real* backGrad);
+    real* backGrad, const int outStride);
 
 /**
  * @brief   Cross-map-respose normalize forward.

paddle/cuda/include/stub/hl_cnn_stub.h

@@ -44,7 +44,8 @@ inline void hl_maxpool_forward(
     const int pooledH, const int pooledW,
     const int sizeX, const int sizeY,
     const int strideH, const int strideW,
-    const int paddingH, const int paddingW, real* tgtData) {}
+    const int paddingH, const int paddingW,
+    real* tgtData, const int tgtStride) {}
 
 inline void hl_maxpool_backward(
     const int frameCnt, const real* inputData,
@@ -56,7 +57,7 @@ inline void hl_maxpool_backward(
     const int strideH, const int strideW,
     const int paddingH, const int paddingW,
     real scaleA, real scaleB,
-    real* targetGrad) {}
+    real* targetGrad, const int outStride) {}
 
 inline void hl_avgpool_forward(
     const int frameCnt, const real* inputData,
@@ -65,7 +66,8 @@ inline void hl_avgpool_forward(
     const int pooledH, const int pooledW,
     const int sizeX, const int sizeY,
     const int strideH, const int strideW,
-    const int paddingH, const int paddingW, real* tgtData) {}
+    const int paddingH, const int paddingW,
+    real* tgtData, const int tgtStride) {}
 
 inline void hl_avgpool_backward(
     const int frameCnt, const real* outGrad,
@@ -76,7 +78,7 @@ inline void hl_avgpool_backward(
     const int strideH, const int strideW,
     int paddingH, int paddingW,
     real scaleA, real scaleB,
-    real* backGrad) {}
+    real* backGrad, const int outStride) {}
 
 inline void hl_CMRNorm_forward(
     size_t frameCnt, const real* in, real* scale, real* out,

paddle/cuda/src/hl_cuda_cnn.cu

@@ -152,7 +152,7 @@ __global__ void KeMaxPoolForward(const int nthreads, const real* inputData,
                                  const int ksizeW, const int ksizeH,
                                  const int strideH, const int strideW,
                                  const int offsetH, const int offsetW,
-                                 real* tgtData) {
+                                 real* tgtData, const int tgtStride) {
   int index =  blockIdx.x * blockDim.x + threadIdx.x;
   if (index < nthreads) {
     int pw = index % pooledW;
@@ -173,7 +173,9 @@ __global__ void KeMaxPoolForward(const int nthreads, const real* inputData,
           maxval = inputData[h * width + w];
       }
     }
-    tgtData[index] = maxval;
+    int tgtIndex = index % (pooledW * pooledH * channels) +
+        frameNum * tgtStride;
+    tgtData[tgtIndex] = maxval;
   }
 }
 
@@ -184,7 +186,7 @@ void hl_maxpool_forward(const int frameCnt, const real* inputData,
                         const int sizeX, const int sizeY,
                         const int strideH, const int strideW,
                         const int paddingH, const int paddingW,
-                        real* tgtData) {
+                        real* tgtData, const int tgtStride) {
 
   int num_kernels = pooledH * pooledW * channels * frameCnt;
   int blocks = (num_kernels + 1024 - 1) / 1024;
@@ -194,7 +196,7 @@ void hl_maxpool_forward(const int frameCnt, const real* inputData,
   KeMaxPoolForward<<< grid, threads, 0, STREAM_DEFAULT >>>
            (num_kernels, inputData, channels, height, width,
            pooledH, pooledW, sizeX, sizeY, strideH, strideW,
-           paddingH, paddingW, tgtData);
+           paddingH, paddingW, tgtData, tgtStride);
   CHECK_SYNC("hl_maxpool_forward failed");
 }
 
@@ -207,7 +209,7 @@ __global__ void KeMaxPoolBackward(const int nthreads, const real* inputData,
                                   const int strideH, const int strideW,
                                   const int padH, const int padW,
                                   real scaleA, real scaleB,
-                                  real* targetGrad) {
+                                  real* targetGrad, const int outStride) {
   int index = blockIdx.x  * blockDim.x + threadIdx.x;
   if (index < nthreads) {
     // find out the local index
@@ -223,8 +225,8 @@ __global__ void KeMaxPoolBackward(const int nthreads, const real* inputData,
     int pwend = offsetW >= 0 ? min(offsetW / strideW + 1, pooledW) : 0;
     real gradient = 0;
     real input = inputData[index];
-    outData += (frameNum * channels + offsetC) * pooledH * pooledW;
-    outGrad += (frameNum * channels + offsetC) * pooledH * pooledW;
+    outData += (frameNum * outStride + offsetC * pooledH * pooledW);
+    outGrad += (frameNum * outStride + offsetC * pooledH * pooledW);
     for (int ph = phstart; ph < phend; ++ph) {
       for (int pw = pwstart; pw < pwend; ++pw) {
         if (input == outData[ph * pooledW + pw]) {
@@ -246,7 +248,7 @@ void hl_maxpool_backward(const int frameCnt, const real* inputData,
                         const int strideH, const int strideW,
                         const int paddingH, const int paddingW,
                         real scaleA, real scaleB,
-                        real* targetGrad) {
+                        real* targetGrad, const int outStride) {
 
   int num_kernels = height * width * channels * frameCnt;
   int blocks = (num_kernels + 1024 - 1) / 1024;
@@ -257,7 +259,7 @@ void hl_maxpool_backward(const int frameCnt, const real* inputData,
            strideH, strideW,
            paddingH, paddingW,
            scaleA, scaleB,
-           targetGrad);
+           targetGrad, outStride);
   CHECK_SYNC("hl_maxpool_backward");
 }
 
@@ -268,7 +270,7 @@ __global__ void KeAvgPoolForward(const int nthreads, const real* inputData,
                                  const int sizeX, const int sizeY,
                                  const int strideH, const int strideW,
                                  const int padH, const int padW,
-                                 real* tgtData) {
+                                 real* tgtData, const int tgtStride) {
   int index = blockIdx.x * blockDim.x + threadIdx.x;
   if (index < nthreads) {
     int pw = index % pooledW;
@@ -293,7 +295,9 @@ __global__ void KeAvgPoolForward(const int nthreads, const real* inputData,
         aveval += inputData[h * width + w];
       }
     }
-    tgtData[index] = aveval / pool_size;
+    int tgtIndex = index % (pooledW * pooledH * channels) +
+        frameNum * tgtStride;
+    tgtData[tgtIndex] = aveval / pool_size;
   }
 }
 
@@ -303,14 +307,15 @@ void hl_avgpool_forward(const int frameCnt, const real* inputData,
                         const int pooledH, const int pooledW,
                         const int sizeX, const int sizeY,
                         const int strideH, const int strideW,
-                        const int paddingH, const int paddingW, real* tgtData) {
+                        const int paddingH, const int paddingW, 
+                        real* tgtData, const int tgtStride) {
   int num_kernels = pooledH * pooledW * channels * frameCnt;
   int blocks = (num_kernels + 1024 - 1) / 1024;
   KeAvgPoolForward<<< blocks, 1024, 0, STREAM_DEFAULT >>>
            (num_kernels, inputData, channels,
            height, width, pooledH, pooledW,
            sizeX, sizeY, strideH, strideW,
-           paddingH, paddingW, tgtData);
+           paddingH, paddingW, tgtData, tgtStride);
   CHECK_SYNC("hl_avgpool_forward failed");
 }
 
@@ -322,7 +327,7 @@ __global__ void KeAvgPoolBackward(const int nthreads, const real* outGrad,
                                   const int strideH, const int strideW,
                                   const int padH, const int padW,
                                   real scaleA, real scaleB,
-                                  real* tgtGrad) {
+                                  real* tgtGrad, const int outStride) {
   int index = blockIdx.x * blockDim.x + threadIdx.x;
   if (index < nthreads) {
     int offsetW = index % width + padW;
@@ -335,7 +340,8 @@ __global__ void KeAvgPoolBackward(const int nthreads, const real* outGrad,
     int phend = offsetH >= 0 ? min(offsetH / strideH + 1, pooledH) : 0;
     int pwend = offsetW >= 0 ? min(offsetW / strideW + 1, pooledW) : 0;
     real gradient = 0;
-    outGrad += (frameNum * channels + offsetC) * pooledH * pooledW;
+    outGrad += (frameNum * outStride + offsetC * pooledH * pooledW);
+
 
     for (int ph = phstart; ph < phend; ++ph) {
       for (int pw = pwstart; pw < pwend; ++pw) {
@@ -360,7 +366,7 @@ void hl_avgpool_backward(const int frameCnt, const real* outGrad,
                          const int strideH, const int strideW,
                          const int paddingH, const int paddingW,
                          real scaleA, real scaleB,
-                         real* backGrad) {
+                         real* backGrad, const int outStride) {
   int num_kernels = height * width * channels * frameCnt;
   int blocks = (num_kernels + 1024 - 1) / 1024;
 
@@ -370,7 +376,7 @@ void hl_avgpool_backward(const int frameCnt, const real* outGrad,
            strideH, strideW,
            paddingH, paddingW,
            scaleA, scaleB,
-           backGrad);
+           backGrad, outStride);
   CHECK_SYNC("hl_avgpool_backward failed");
 }
 

paddle/gserver/layers/PoolLayer.cpp

@@ -52,10 +52,8 @@ bool PoolLayer::init(const LayerMap& layerMap,
 Layer* PoolLayer::create(const LayerConfig& config) {
   CHECK_EQ(config.inputs_size(), 1);
   const std::string& pool = config.inputs(0).pool_conf().pool_type();
-  if (pool == "max-projection") {
-    return new MaxPoolProjectionLayer(config);
-  } else if (pool == "avg-projection") {
-    return new AvgPoolProjectionLayer(config);
+  if (pool == "max-projection" || pool == "avg-projection") {
+    return new PoolProjectionLayer(config);
 #ifndef PADDLE_ONLY_CPU
   } else if (CudnnPoolLayer::typeCheck(pool)) {
     return new CudnnPoolLayer(config);

paddle/gserver/layers/PoolProjection.cpp

+/* Copyright (c) 2016 Baidu, Inc. 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 "PoolProjection.h"
+
+namespace paddle {
+
+REGISTER_PROJECTION_CREATE_FUNC(pool, &PoolProjection::create);
+
+PoolProjection::PoolProjection(const ProjectionConfig& config,
+                               ParameterPtr parameter, bool useGpu)
+    : Projection(config, parameter, useGpu) {
+  const PoolConfig& conf = config_.pool_conf();
+  poolType_ = conf.pool_type();
+  channels_ = conf.channels();
+  sizeX_ = conf.size_x();
+  stride_ = conf.stride();
+  outputX_ = conf.output_x();
+  imgSize_ = conf.img_size();
+  confPadding_ = conf.padding();
+
+  sizeY_ = conf.has_size_y() ? conf.size_y() : conf.size_x();
+  imgSizeY_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
+  strideY_ = conf.has_stride_y() ? conf.stride_y() : conf.stride();
+  confPaddingY_ = conf.has_padding_y() ? conf.padding_y() : conf.padding();
+  outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x();
+}
+
+size_t PoolProjection::getSize() {
+  imgSizeY_ = in_->getFrameHeight();
+  imgSize_ = in_->getFrameWidth();
+  const PoolConfig& conf = config_.pool_conf();
+  if (imgSizeY_ == 0) {
+    imgSizeY_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
+  }
+  if (imgSize_ == 0) {
+    imgSize_ = conf.img_size();
+  }
+  outputY_ = outputSize(imgSizeY_, sizeY_, confPaddingY_, strideY_,
+                        /* caffeMode */ false);
+  outputX_ = outputSize(imgSize_, sizeX_, confPadding_, stride_,
+                        /* caffeMode */ false);
+
+  const_cast<Argument*>(out_)->setFrameHeight(outputY_);
+  const_cast<Argument*>(out_)->setFrameWidth(outputX_);
+
+  return outputY_ * outputX_ * channels_;
+}
+
+PoolProjection* PoolProjection::create(const ProjectionConfig& config,
+                                       ParameterPtr parameter, bool useGpu) {
+  const std::string& pool = config.pool_conf().pool_type();
+  if (pool == "max-projection") {
+    return new MaxPoolProjection(config, parameter, useGpu);
+  } else if (pool == "avg-projection") {
+    return new AvgPoolProjection(config, parameter, useGpu);
+  } else {
+    LOG(FATAL) << "Unknown pool type: " << pool;
+    return nullptr;
+  }
+}
+
+void MaxPoolProjection::forward() {
+  size_t width = getSize();
+  CHECK_EQ(width, out_->value->getWidth());
+  MatrixPtr inputV = in_->value;
+  MatrixPtr outV = out_->value;
+  outV->maxPoolForward(*inputV, imgSizeY_, imgSize_, channels_, sizeX_, sizeY_,
+                       strideY_, stride_, outputY_, outputX_, confPaddingY_,
+                       confPadding_);
+}
+
+void MaxPoolProjection::backward(const UpdateCallback& callback) {
+  (void)callback;
+  MatrixPtr outGrad = out_->grad;
+  MatrixPtr inputV = in_->value;
+  MatrixPtr outV = out_->value;
+  MatrixPtr inputGrad = in_->grad;
+
+  if (NULL == inputGrad) {
+    return;
+  }
+  inputGrad->maxPoolBackward(*inputV, imgSizeY_, imgSize_, *outGrad, *outV,
+                             sizeX_, sizeY_, strideY_, stride_, outputY_,
+                             outputX_, 1, 1, confPaddingY_, confPadding_);
+}
+
+void AvgPoolProjection::forward() {
+  size_t width = getSize();
+  CHECK_EQ(width, out_->value->getWidth());
+  MatrixPtr inputV = in_->value;
+  MatrixPtr outV = out_->value;
+  outV->avgPoolForward(*inputV, imgSizeY_, imgSize_, channels_, sizeX_, sizeY_,
+                       strideY_, stride_, outputY_, outputX_, confPaddingY_,
+                       confPadding_);
+}
+
+void AvgPoolProjection::backward(const UpdateCallback& callback) {
+  (void)callback;
+
+  MatrixPtr outputGrad = out_->grad;
+  MatrixPtr inputGrad = in_->grad;
+
+  if (NULL == inputGrad) {
+    return;
+  }
+
+  inputGrad->avgPoolBackward(*outputGrad, imgSizeY_, imgSize_, sizeX_, sizeY_,
+                             strideY_, stride_, outputY_, outputX_, 1, 1,
+                             confPaddingY_, confPadding_);
+}
+}  // namespace paddle

paddle/gserver/layers/PoolProjection.h

+/* Copyright (c) 2016 Baidu, Inc. 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 "Projection.h"
+#include "paddle/math/MathUtils.h"
+
+namespace paddle {
+
+class PoolProjection : public Projection {
+protected:
+  size_t imgSizeY_, imgSize_;
+  size_t outputY_, outputX_;
+  size_t strideY_, stride_;
+  size_t sizeY_, sizeX_;
+  int confPaddingY_, confPadding_;
+  size_t channels_;
+  std::string poolType_;
+
+public:
+  PoolProjection(const ProjectionConfig& config, ParameterPtr parameter,
+                 bool useGpu);
+
+  static PoolProjection* create(const ProjectionConfig& config,
+                                ParameterPtr parameter, bool useGpu);
+
+  const std::string& getPoolType() const { return poolType_; }
+
+  size_t getSize();
+};
+
+class MaxPoolProjection : public PoolProjection {
+public:
+  MaxPoolProjection(const ProjectionConfig& config, ParameterPtr parameter,
+                    bool useGpu)
+      : PoolProjection(config, parameter, useGpu) {}
+
+  virtual void forward();
+  virtual void backward(const UpdateCallback& callback = nullptr);
+};
+
+class AvgPoolProjection : public PoolProjection {
+public:
+  AvgPoolProjection(const ProjectionConfig& config, ParameterPtr parameter,
+                    bool useGpu)
+      : PoolProjection(config, parameter, useGpu) {}
+
+  virtual void forward();
+  virtual void backward(const UpdateCallback& callback = nullptr);
+};
+}  // namespace paddle

paddle/gserver/layers/PoolProjectionLayer.cpp

@@ -18,6 +18,7 @@ limitations under the License. */
 
 namespace paddle {
 
+
 size_t PoolProjectionLayer::getSize() {
   CHECK_EQ(inputLayers_.size(), 1UL);
   size_t layerSize = 0;
@@ -37,74 +38,23 @@ size_t PoolProjectionLayer::getSize() {
 
   layerSize = outputH_ * outputW_ * channels_;
 
-  getOutput().setFrameHeight(outputH_);
-  getOutput().setFrameWidth(outputW_);
   return layerSize;
 }
 
-void MaxPoolProjectionLayer::forward(PassType passType) {
-  Layer::forward(passType);
-
-  /* malloc memory for the output_ if necessary */
-  /* note: one sample correspond to one ROW */
-  MatrixPtr input = getInputValue(0);
-  int batchSize = input->getHeight();
-  int size = getSize();
-  resetOutput(batchSize, size);
-
-  MatrixPtr outV = getOutputValue();
-
-  outV->maxPoolForward(*input, imgSizeH_, imgSizeW_, channels_, sizeX_, sizeY_,
-                       strideY_, stride_, outputH_, outputW_, confPaddingY_,
-                       confPadding_);
-}
-
-void MaxPoolProjectionLayer::backward(const UpdateCallback& callback) {
-  (void)callback;
-
-  if (NULL == getInputGrad(0)) {
-    return;
-  }
-
-  /* Do derivation */
-  MatrixPtr outGrad = getOutputGrad();
-  MatrixPtr inputV = getInputValue(0);
-  MatrixPtr outV = getOutputValue();
-  MatrixPtr inputGrad = getInputGrad(0);
-
-  inputGrad->maxPoolBackward(*inputV, imgSizeH_, imgSizeW_, *outGrad, *outV,
-                             sizeX_, sizeY_, strideY_, stride_, outputH_,
-                             outputW_, 1, 1, confPaddingY_, confPadding_);
-}
-
-void AvgPoolProjectionLayer::forward(PassType passType) {
+void PoolProjectionLayer::forward(PassType passType) {
   Layer::forward(passType);
-
-  /* malloc memory for the output_ if necessary */
-  /* note: one sample correspond to one ROW */
-  MatrixPtr input = getInputValue(0);
-  int batchSize = input->getHeight();
+  const Argument& in = getInput(0);
+  int batchSize = in.value->getHeight();
   int size = getSize();
   resetOutput(batchSize, size);
-
-  MatrixPtr outV = getOutputValue();
-
-  outV->avgPoolForward(*input, imgSizeH_, imgSizeW_, channels_, sizeX_, sizeY_,
-                       strideY_, stride_, outputH_, outputW_, confPaddingY_,
-                       confPadding_);
+  poolProjection_->forward(&in, &output_, passType);
 }
 
-void AvgPoolProjectionLayer::backward(const UpdateCallback& callback) {
+void PoolProjectionLayer::backward(const UpdateCallback& callback) {
   (void)callback;
-
   if (NULL == getInputGrad(0)) {
     return;
   }
-  /* Do derivation */
-  MatrixPtr outputGrad = getOutputGrad();
-  MatrixPtr inputGrad = getInputGrad(0);
-  inputGrad->avgPoolBackward(*outputGrad, imgSizeH_, imgSizeW_, sizeX_, sizeY_,
-                             strideY_, stride_, outputH_, outputW_, 1, 1,
-                             confPaddingY_, confPadding_);
+  poolProjection_->backward(callback);
 }
 }  // namespace paddle

paddle/gserver/layers/PoolProjectionLayer.h

@@ -12,12 +12,12 @@ 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 "PoolLayer.h"
+#include "PoolProjection.h"
 #include "paddle/math/Matrix.h"
-#include <vector>
 
 namespace paddle {
 /**
@@ -27,33 +27,18 @@ class PoolProjectionLayer : public PoolLayer {
 protected:
   size_t imgSizeH_, imgSizeW_;
   size_t outputH_, outputW_;
+  std::unique_ptr<PoolProjection> poolProjection_;
+  ProjectionConfig projectionConfig_;
 
 public:
-  size_t getSize();
-  explicit PoolProjectionLayer(const LayerConfig& config) : PoolLayer(config) {}
-};
-/**
- * @brief A layer for max pooling
- */
-class MaxPoolProjectionLayer : public PoolProjectionLayer {
-public:
-  explicit MaxPoolProjectionLayer(const LayerConfig& config)
-      : PoolProjectionLayer(config) {}
-
-  ~MaxPoolProjectionLayer() {}
+  explicit PoolProjectionLayer(const LayerConfig& config) : PoolLayer(config) {
+    PoolConfig* conf = projectionConfig_.mutable_pool_conf();
+    *conf = config_.inputs(0).pool_conf();
+    poolProjection_.reset(
+        PoolProjection::create(projectionConfig_, nullptr, useGpu_));
+  }
 
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
-};
-/**
- * @brief A layer for average pooling
- */
-class AvgPoolProjectionLayer : public PoolProjectionLayer {
-public:
-  explicit AvgPoolProjectionLayer(const LayerConfig& config)
-      : PoolProjectionLayer(config) {}
-
-  ~AvgPoolProjectionLayer() {}
+  size_t getSize();
 
   virtual void forward(PassType passType);
   virtual void backward(const UpdateCallback& callback = nullptr);

paddle/gserver/layers/Projection.h

@@ -12,12 +12,11 @@ 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/parameter/Parameter.h"
-#include "ModelConfig.pb.h"
 #include "Layer.h"
+#include "ModelConfig.pb.h"
+#include "paddle/parameter/Parameter.h"
 
 namespace paddle {
 
@@ -28,6 +27,11 @@ namespace paddle {
     Projection::registrar_.registerClass<__class_name>(#__type_name); \
   })
 
+#define REGISTER_PROJECTION_CREATE_FUNC(__type_name, createFunction)    \
+  static InitFunction __reg_type_##__type_name([]() {                   \
+    Projection::registrar_.registerClass(#__type_name, createFunction); \
+  })
+
 /**
  * A projection takes one Argument as input, calculate the result and add it
  * to output Argument.
@@ -50,7 +54,8 @@ public:
       registrar_;
 
   /**
-   * Forward propagation. If backward() will be called, in and out must be kept valid until then.
+   * Forward propagation. If backward() will be called, in and out must be kept
+   * valid until then.
    * @param in input of projection
    * @param out output of projection
    * @param passType PASS_TRAIN of PASS_TEST

paddle/gserver/layers/SpatialPyramidPoolLayer.cpp

+/* Copyright (c) 2016 Baidu, Inc. 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 "SpatialPyramidPoolLayer.h"
+
+namespace paddle {
+
+REGISTER_LAYER(spp, SpatialPyramidPoolLayer);
+
+ProjectionConfig SpatialPyramidPoolLayer::getConfig(size_t imgSizeW,
+                                                    size_t imgSizeH,
+                                                    size_t channels,
+                                                    size_t pyramidLevel,
+                                                    std::string& poolType) {
+  ProjectionConfig config;
+  config.set_type("pool");
+  PoolConfig* conf = config.mutable_pool_conf();
+  conf->set_channels(channels);
+  conf->set_img_size(imgSizeW);
+  conf->set_img_size_y(imgSizeH);
+  conf->set_pool_type(poolType);
+
+  int numBins = std::pow(2, pyramidLevel);
+
+  int sizeH = std::ceil(imgSizeH / static_cast<double>(numBins));
+  int paddingH = (sizeH * numBins - imgSizeH + 1) / 2;
+  int outSizeH = outputSize(imgSizeH, sizeH, paddingH, sizeH, true);
+
+  int sizeW = std::ceil(imgSizeW / static_cast<double>(numBins));
+  int paddingW = (sizeW * numBins - imgSizeW + 1) / 2;
+  int outSizeW = outputSize(imgSizeW, sizeW, paddingW, sizeW, true);
+
+  conf->set_stride(sizeW);
+  conf->set_stride_y(sizeH);
+  conf->set_size_x(sizeW);
+  conf->set_size_y(sizeH);
+  conf->set_padding(paddingW);
+  conf->set_padding_y(paddingH);
+  conf->set_output_x(outSizeW);
+  conf->set_output_y(outSizeH);
+  config.set_output_size(outSizeH * outSizeW * channels);
+  return config;
+}
+
+size_t SpatialPyramidPoolLayer::getSize() {
+  CHECK_EQ(inputLayers_.size(), 1UL);
+  size_t layerSize = 0;
+  const SppConfig& sppConf = config_.inputs(0).spp_conf();
+  imgSizeH_ = inputLayers_[0]->getOutput().getFrameHeight();
+  imgSizeW_ = inputLayers_[0]->getOutput().getFrameWidth();
+  if (imgSizeH_ == 0) {
+    imgSizeH_ = sppConf.has_img_size_y() ? sppConf.img_size_y() : imgSizeW_;
+  }
+  if (imgSizeW_ == 0) {
+    imgSizeW_ = sppConf.img_size();
+  }
+
+  size_t outputH = 1;
+  size_t outputW = (std::pow(4, pyramidHeight_) - 1) / (4 - 1);
+
+  layerSize = outputH * outputW * channels_;
+  return layerSize;
+}
+
+bool SpatialPyramidPoolLayer::init(const LayerMap& layerMap,
+                                   const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+  CHECK_EQ(config_.inputs_size(), 1);
+
+  const SppConfig& sppConf = config_.inputs(0).spp_conf();
+  pyramidHeight_ = sppConf.pyramid_height();
+  poolType_ = sppConf.pool_type();
+
+  channels_ = sppConf.channels();
+  imgSizeW_ = sppConf.img_size();
+  imgSizeH_ = sppConf.has_img_size_y() ? sppConf.img_size_y() : imgSizeW_;
+  poolProjections_.reserve(pyramidHeight_);
+  projCol_.reserve(pyramidHeight_);
+  projOutput_.resize(pyramidHeight_);
+
+  size_t startCol = 0;
+  size_t endCol = 0;
+  for (size_t i = 0; i < pyramidHeight_; i++) {
+    poolProjections_.emplace_back(PoolProjection::create(
+        getConfig(imgSizeW_, imgSizeH_, channels_, i, poolType_), nullptr,
+        useGpu_));
+    endCol += poolProjections_[i]->getOutputSize();
+    projCol_.push_back(std::make_pair(startCol, endCol));
+    startCol = endCol;
+  }
+  CHECK_EQ(endCol, getSize());
+  return true;
+}
+
+void SpatialPyramidPoolLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  int batchSize = getInput(0).getBatchSize();
+  resetOutput(batchSize, getSize());
+  for (size_t i = 0; i < pyramidHeight_; i++) {
+    size_t startCol = projCol_[i].first;
+    size_t endCol = projCol_[i].second;
+    projOutput_[i].value = output_.value->subColMatrix(startCol, endCol);
+    projOutput_[i].grad = output_.grad->subColMatrix(startCol, endCol);
+  }
+  for (size_t i = 0; i < pyramidHeight_; i++) {
+    poolProjections_[i]->forward(&getInput(0), &projOutput_[i], passType);
+  }
+}
+
+void SpatialPyramidPoolLayer::backward(const UpdateCallback& callback) {
+  for (size_t i = 0; i < pyramidHeight_; i++) {
+    if (poolProjections_[i]) {
+      poolProjections_[i]->backward(callback);
+    }
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/SpatialPyramidPoolLayer.h

+/* Copyright (c) 2016 Baidu, Inc. 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"
+#include "PoolProjection.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/utils/Logging.h"
+
+namespace paddle {
+/**
+ * @brief A layer for spatial pyramid pooling on the input image by taking
+ * the max, average, etc. within regions, so that the result vector of
+ * different sized images are of the same size.
+ * 
+ * The config file api is spp_layer.
+ */
+
+class SpatialPyramidPoolLayer : public Layer {
+protected:
+  size_t channels_;
+  size_t imgSizeW_;
+  size_t imgSizeH_;
+  size_t pyramidHeight_;
+  std::string poolType_;
+
+  std::vector<std::unique_ptr<PoolProjection>> poolProjections_;
+  std::vector<Argument> projOutput_;
+  std::vector<std::pair<size_t, size_t>> projCol_;
+
+public:
+  explicit SpatialPyramidPoolLayer(const LayerConfig& config) : Layer(config) {}
+
+  ~SpatialPyramidPoolLayer() {}
+
+  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+
+  ProjectionConfig getConfig(size_t sizeX_, size_t sizeY_, size_t channels,
+                             size_t pyamidLevel_, std::string& poolType_);
+  size_t getSize();
+
+  virtual void forward(PassType passType);
+  virtual void backward(const UpdateCallback& callback = nullptr);
+};
+}  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -13,15 +13,15 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <gtest/gtest.h>
-#include <vector>
 #include <string>
-#include "paddle/gserver/layers/DataLayer.h"
+#include <vector>
 #include "ModelConfig.pb.h"
+#include "paddle/gserver/layers/DataLayer.h"
 #include "paddle/trainer/Trainer.h"
 #include "paddle/math/MathUtils.h"
 
-#include "TestUtil.h"
 #include "LayerGradUtil.h"
+#include "TestUtil.h"
 
 using namespace paddle;  // NOLINT
 using namespace std;     // NOLINT
@@ -981,6 +981,32 @@ TEST(Layer, PoolLayer) {
 #endif
 }
 
+void testSppLayer(const string& poolType, const int pyramidHeight, bool trans,
+                  bool useGpu) {
+  TestConfig config;
+  config.layerConfig.set_type("spp");
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 3200, 0});
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  SppConfig* sppConfig = input->mutable_spp_conf();
+  sppConfig->set_pool_type(poolType);
+  sppConfig->set_pyramid_height(pyramidHeight);
+  sppConfig->set_channels(16);
+  sppConfig->set_img_size(10);
+  sppConfig->set_img_size_y(20);
+  int outputSize = (std::pow(4, sppConfig->pyramid_height()) - 1) / (4 - 1);
+  config.layerConfig.set_size(outputSize * sppConfig->channels());
+  testLayerGrad(config, "spp", 100, trans, useGpu);
+}
+
+TEST(Layer, SpatialPyramidPoolLayer) {
+  for (auto useGpu : {false, true}) {
+    for (auto pyramidHeight : {1, 2, 3}) {
+      testSppLayer("avg-projection", pyramidHeight, false, useGpu);
+      testSppLayer("max-projection", pyramidHeight, false, useGpu);
+    }
+  }
+}
+
 TEST(Layer, rankCostLayer) {
   TestConfig config;
   config.layerConfig.set_type("rank-cost");

proto/ModelConfig.proto.m4

@@ -120,6 +120,14 @@ message PoolConfig {
   optional uint32 padding_y = 13 [default = 0];
 }
 
+message SppConfig {
+  required string pool_type = 1;
+  required uint32 pyramid_height = 2;
+  required uint32 channels = 3;
+  required uint32 img_size = 4;
+  optional uint32 img_size_y = 5;
+}
+
 message NormConfig {
   // rnorm or cmrnorm
   required string norm_type = 1;
@@ -196,6 +204,9 @@ message ProjectionConfig {
 
   // For IdentityOffsetProjection
   optional uint64 offset = 11 [default = 0];
+
+  // For pool
+  optional PoolConfig pool_conf = 12;
 }
 
 message OperatorConfig {
@@ -245,6 +256,7 @@ message LayerInputConfig {
   optional string input_layer_argument = 9;
   optional BilinearInterpConfig bilinear_interp_conf = 10;
   optional MaxOutConfig maxout_conf = 11;
+  optional SppConfig spp_conf = 12;
 }
 
 message LayerConfig {

python/paddle/trainer/config_parser.py

@@ -471,6 +471,7 @@ class Input(Cfg):
             image=None,
             block_expand=None,
             maxout=None,
+            spp=None,
             format=None,
             nnz=None,
             is_static=None,
@@ -671,7 +672,6 @@ class ConvProjection(Projection):
     def calc_parameter_dims(self, input_size, output_size):
         return None
 
-
 # Define a operator for mixed layer
 @config_class
 class Operator(Cfg):
@@ -796,6 +796,17 @@ class Pool(Cfg):
             padding_y = None):
         self.add_keys(locals())
         
+# please refer to the comments in proto/ModelConfig.proto
+@config_class
+class SpatialPyramidPool(Cfg):
+    def __init__(
+            self,
+            pool_type,
+            pyramid_height,
+            channels,
+            img_width = None):
+        self.add_keys(locals())
+
 # please refer to the comments in proto/ModelConfig.proto
 @config_class
 class Norm(Cfg):
@@ -1081,6 +1092,22 @@ def parse_pool(pool, input_layer_name, pool_conf):
         pool_conf.output_y = cnn_output_size(pool_conf.img_size_y, pool_conf.size_y,
                                              pool_conf.padding_y, pool_conf.stride_y, False)
 
+def parse_spp(spp, input_layer_name, spp_conf):
+    spp_conf.pool_type = spp.pool_type
+    config_assert(spp.pool_type in ['max-projection', 'avg-projection'],
+                  "pool-type %s is not in " "['max-projection', 'avg-projection']"
+                  % spp.pool_type)
+    spp_conf.pyramid_height = spp.pyramid_height
+    spp_conf.channels = spp.channels
+
+    img_pixels = g_layer_map[input_layer_name].size / spp_conf.channels
+
+    spp_conf.img_size = default(spp.img_width, int(img_pixels ** 0.5))
+    spp_conf.img_size_y = img_pixels / spp_conf.img_size
+    config_assert(spp_conf.img_size * spp_conf.img_size_y == img_pixels,
+                  "Incorrect input image size %d for input image pixels %d"
+                  % (spp_conf.img_size, img_pixels))
+
 def parse_image(image, input_layer_name, image_conf):
     image_conf.channels = image.channels
     image_pixels = g_layer_map[input_layer_name].size / image_conf.channels
@@ -1756,6 +1783,25 @@ class PoolLayer(LayerBase):
                 name, pool_conf.output_y, pool_conf.output_x))
             self.set_layer_size((pool_conf.output_x * pool_conf.output_y) * pool_conf.channels)
 
+@config_layer('spp')
+class SpatialPyramidPoolLayer(LayerBase):
+    def __init__(
+            self,
+            name,
+            inputs,
+            device=None):
+        super(SpatialPyramidPoolLayer, self).__init__(name, 'spp', 0, inputs=inputs, device=device)
+        for input_index in xrange(len(self.inputs)):
+            input_layer = self.get_input_layer(input_index)
+            parse_spp(
+                self.inputs[input_index].spp,
+                input_layer.name,
+                self.config.inputs[input_index].spp_conf)
+            spp_conf = self.config.inputs[input_index].spp_conf
+            output_size = (pow(4, spp_conf.pyramid_height) - 1) / (4 - 1)
+            print("output size for %s is %d " % (name, output_size))
+            self.set_layer_size(output_size * spp_conf.channels)
+
 @config_layer('batch_norm')
 class BatchNormLayer(LayerBase):
     layer_type = 'batch_norm'

python/paddle/trainer_config_helpers/layers.py

@@ -56,7 +56,8 @@ __all__ = ["full_matrix_projection", "AggregateLevel", "ExpandLevel",
            'multi_binary_label_cross_entropy', 'sum_cost',
            'rank_cost', 'lambda_cost', 'huber_cost',
            'block_expand_layer',
-           'maxout_layer', 'out_prod_layer', 'print_layer'
+           'maxout_layer', 'out_prod_layer', 'print_layer', 
+           'spp_layer', 
            ]
 
 
@@ -115,6 +116,7 @@ class LayerType(object):
     LINEAR_COMBINATION_LAYER = "convex_comb"
     BLOCK_EXPAND = "blockexpand"
     MAXOUT = "maxout"
+    SPP_LAYER = "spp"
 
     PRINT_LAYER = "print"
 
@@ -877,6 +879,7 @@ def pooling_layer(input, pooling_type=None, name=None, bias_attr=None,
                        size=input.size)
 
 
+
 @wrap_bias_attr_default()
 @wrap_param_attr_default()
 @wrap_act_default(param_names=['gate_act'],
@@ -1820,6 +1823,62 @@ def img_pool_layer(input, pool_size, name=None,
                        num_filters=num_channels, size=l.config.size)
 
 
+@wrap_name_default("spp")
+@layer_support()
+def spp_layer(input, name=None, num_channels=None, pool_type=None,
+              pyramid_height=None, img_width=None, layer_attr=None):
+    pass
+    """
+    Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition.
+    The details please refer to
+    `Kaiming He's paper <https://arxiv.org/abs/1406.4729>`_.
+
+    :param name: layer name.
+    :type name: basestring
+    :param input: layer's input.
+    :type input: LayerOutput
+    :param num_channels: number of input channel.
+    :type num_channels: int
+    :param pool_type: Pooling type. MaxPooling or AveragePooling. Default is MaxPooling.
+    :type scale: BasePoolingType
+    :param pyramid_height: pyramid height.
+    :type pyramid_height: int
+    :param img_width: the width of input feature map. If it is None, the input feature
+                      map should be square.
+    :type img_width: int|None
+    :param layer_attr: Extra Layer Attribute.
+    :type layer_attr: ExtraLayerAttribute
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+    if num_channels is None:
+        assert input.num_filters is not None
+        num_channels = input.num_filters
+
+    if pool_type is None:
+        pool_type = MaxPooling()
+    elif isinstance(pool_type, AvgPooling):
+        pool_type.name = 'avg'
+
+    type_name = pool_type.name
+    if (isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)):
+        type_name += '-projection'
+
+    Layer(
+        name=name,
+        type=LayerType.SPP_LAYER,
+        inputs=Input(input.name,
+                     spp=SpatialPyramidPool(pool_type=type_name,
+                                            channels=num_channels,
+                                            pyramid_height=pyramid_height,
+                                            img_width=img_width)
+        ),
+        **ExtraLayerAttribute.to_kwargs(layer_attr)
+    )
+    return LayerOutput(name, LayerType.SPP_LAYER, parents=[input], 
+                       num_filters=num_channels)
+
+
 def __img_norm_layer__(name, input, size, norm_type, scale, power,
                        num_channels, blocked, layer_attr):
     if num_channels is None:

python/paddle/trainer_config_helpers/tests/configs/generate_protostr.sh

@@ -11,7 +11,7 @@ test_sequence_pooling test_lstmemory_layer test_grumemory_layer
 last_first_seq test_expand_layer test_ntm_layers test_hsigmoid
 img_layers img_trans_layers util_layers simple_rnn_layers unused_layers test_cost_layers
 test_rnn_group shared_fc shared_lstm test_cost_layers_with_weight
-test_bilinear_interp test_maxout test_bi_grumemory math_ops)
+test_spp_layer test_bilinear_interp test_maxout test_bi_grumemory math_ops)
 
 
 for conf in ${configs[*]}

python/paddle/trainer_config_helpers/tests/configs/protostr/test_spp_layer.protostr

+type: "nn"
+layers {
+  name: "data"
+  type: "data"
+  size: 3200
+  active_type: ""
+}
+layers {
+  name: "__spp_0__"
+  type: "spp"
+  size: 80
+  active_type: ""
+  inputs {
+    input_layer_name: "data"
+    spp_conf {
+      pool_type: "max-projection"
+      pyramid_height: 2
+      channels: 16
+      img_size: 10
+      img_size_y: 20
+    }
+  }
+}
+input_layer_names: "data"
+output_layer_names: "__spp_0__"
+sub_models {
+  name: "root"
+  layer_names: "data"
+  layer_names: "__spp_0__"
+  input_layer_names: "data"
+  output_layer_names: "__spp_0__"
+  is_recurrent_layer_group: false
+}
+

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

+from paddle.trainer_config_helpers import *
+
+settings(
+    batch_size=100,
+    learning_rate=1e-5
+)
+
+data = data_layer(name='data', size=3200)
+
+spp = spp_layer(input=data,
+                pyramid_height=2,
+                num_channels=16,
+                pool_type=MaxPooling(),
+                img_width=10)
+
+outputs(spp)