Merge pull request #1653 from Noplz/normalize-layer

CrossChannelNorm Layer for SSD

doc/api/v2/config/layer.rst

@@ -109,6 +109,12 @@ sum_to_one_norm
     :members: sum_to_one_norm
     :noindex:
     
+cross_channel_norm
+------------------
+..  automodule:: paddle.v2.layer
+    :members: cross_channel_norm
+    :noindex:
+    
 Recurrent Layers
 ================
 

paddle/gserver/layers/CrossChannelNormLayer.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 "Layer.h"
+#include "NormLayer.h"
+#include "paddle/math/BaseMatrix.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+MatrixPtr CrossChannelNormLayer::createSampleMatrix(MatrixPtr data,
+                                                    size_t iter,
+                                                    size_t spatialDim) {
+  return Matrix::create(data->getData() + iter * channels_ * spatialDim,
+                        channels_,
+                        spatialDim,
+                        false,
+                        useGpu_);
+}
+
+MatrixPtr CrossChannelNormLayer::createSpatialMatrix(MatrixPtr data,
+                                                     size_t iter,
+                                                     size_t spatialDim) {
+  return Matrix::create(
+      data->getData() + iter * spatialDim, 1, spatialDim, false, useGpu_);
+}
+
+void CrossChannelNormLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  MatrixPtr inV = getInputValue(0);
+
+  size_t batchSize = inV->getHeight();
+  size_t dataDim = inV->getWidth();
+  CHECK_EQ(getSize(), dataDim);
+
+  reserveOutput(batchSize, dataDim);
+  MatrixPtr outV = getOutputValue();
+  size_t spatialDim = dataDim / channels_;
+
+  Matrix::resizeOrCreate(dataBuffer_, batchSize, dataDim, false, useGpu_);
+  Matrix::resizeOrCreate(spatialBuffer_, 1, spatialDim, false, useGpu_);
+  Matrix::resizeOrCreate(normBuffer_, batchSize, spatialDim, false, useGpu_);
+  normBuffer_->zeroMem();
+  // add eps to avoid overflow
+  normBuffer_->addScalar(*normBuffer_, 1e-6);
+  inV->square2(*dataBuffer_);
+  for (size_t i = 0; i < batchSize; i++) {
+    const MatrixPtr inVTmp = createSampleMatrix(inV, i, spatialDim);
+    const MatrixPtr dataTmp = createSampleMatrix(dataBuffer_, i, spatialDim);
+    MatrixPtr outVTmp = createSampleMatrix(outV, i, spatialDim);
+    MatrixPtr normTmp = createSpatialMatrix(normBuffer_, i, spatialDim);
+
+    // compute norm.
+    spatialBuffer_->sumCols(*dataTmp, 1, 0);
+    spatialBuffer_->sqrt2(*spatialBuffer_);
+    normTmp->copyFrom(*spatialBuffer_);
+    outVTmp->copyFrom(*inVTmp);
+    outVTmp->divRowVector(*spatialBuffer_);
+    // scale the layer.
+    outVTmp->mulColVector(*scale_->getW());
+  }
+}
+
+void CrossChannelNormLayer::backward(const UpdateCallback& callback) {
+  MatrixPtr inG = getInputGrad(0);
+  MatrixPtr inV = getInputValue(0);
+  MatrixPtr outG = getOutputGrad();
+  MatrixPtr outV = getOutputValue();
+
+  size_t batchSize = inG->getHeight();
+  size_t dataDim = inG->getWidth();
+  size_t spatialDim = dataDim / channels_;
+
+  dataBuffer_->dotMul(*outG, *outV);
+  Matrix::resizeOrCreate(scaleDiff_, channels_, 1, false, useGpu_);
+  Matrix::resizeOrCreate(channelBuffer_, channels_, 1, false, useGpu_);
+  Matrix::resizeOrCreate(sampleBuffer_, channels_, spatialDim, false, useGpu_);
+  scaleDiff_->zeroMem();
+  for (size_t i = 0; i < batchSize; i++) {
+    MatrixPtr outGTmp = createSampleMatrix(outG, i, spatialDim);
+    const MatrixPtr dataTmp = createSampleMatrix(dataBuffer_, i, spatialDim);
+    const MatrixPtr inVTmp = createSampleMatrix(inV, i, spatialDim);
+    const MatrixPtr inGTmp = createSampleMatrix(inG, i, spatialDim);
+    const MatrixPtr normTmp = createSpatialMatrix(normBuffer_, i, spatialDim);
+
+    channelBuffer_->sumRows(*dataTmp, 1, 0);
+    channelBuffer_->dotDiv(*channelBuffer_, *(scale_->getW()));
+    // store a / scale[i] in scaleDiff_ temporary
+    scaleDiff_->add(*channelBuffer_, 1.);
+
+    sampleBuffer_->dotMul(*inVTmp, *outGTmp);
+    spatialBuffer_->sumCols(*sampleBuffer_, 1., 1.);
+    // scale the grad
+    inGTmp->copyFrom(*inVTmp);
+    inGTmp->mulRowVector(*spatialBuffer_);
+    // divide by square of norm
+    spatialBuffer_->dotMul(*normTmp, *normTmp);
+    inGTmp->divRowVector(*spatialBuffer_);
+    // subtract
+    inGTmp->add(*outGTmp, -1, 1);
+    // divide by norm
+    inGTmp->divRowVector(*normTmp);
+    // scale the diff
+    inGTmp->mulColVector(*scale_->getW());
+  }
+  // updata scale
+  if (scale_->getWGrad()) scale_->getWGrad()->copyFrom(*scaleDiff_);
+  scale_->getParameterPtr()->incUpdate(callback);
+}
+
+}  // namespace paddle

paddle/gserver/layers/NormLayer.cpp

@@ -26,6 +26,8 @@ Layer* NormLayer::create(const LayerConfig& config) {
     return new ResponseNormLayer(config);
   } else if (norm == "cmrnorm-projection") {
     return new CMRProjectionNormLayer(config);
+  } else if (norm == "cross-channel-norm") {
+    return new CrossChannelNormLayer(config);
   } else {
     LOG(FATAL) << "Unknown norm type: " << norm;
     return nullptr;
@@ -54,4 +56,14 @@ bool ResponseNormLayer::init(const LayerMap& layerMap,
   return true;
 }
 
+bool CrossChannelNormLayer::init(const LayerMap& layerMap,
+                                 const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+  CHECK(parameters_[0]);
+  const NormConfig& conf = config_.inputs(0).norm_conf();
+  channels_ = conf.channels();
+  scale_.reset(new Weight(channels_, 1, parameters_[0]));
+  return true;
+}
+
 }  // namespace paddle

paddle/gserver/layers/NormLayer.h

@@ -65,4 +65,35 @@ public:
   }
 };
 
+/**
+ * This layer applys normalization across the channels of each sample to a
+ * conv layer's output, and scales the output by a group of trainable factors
+ * whose dimensions equal to the number of channels.
+ * - Input: One and only one input layer are accepted.
+ * - Output: The normalized data of the input data.
+ * Reference:
+ *    Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed,
+ *    Cheng-Yang Fu, Alexander C. Berg. SSD: Single Shot MultiBox Detector
+ */
+class CrossChannelNormLayer : public NormLayer {
+public:
+  explicit CrossChannelNormLayer(const LayerConfig& config)
+      : NormLayer(config) {}
+  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  void forward(PassType passType);
+  void backward(const UpdateCallback& callback);
+  MatrixPtr createSampleMatrix(MatrixPtr data, size_t iter, size_t spatialDim);
+  MatrixPtr createSpatialMatrix(MatrixPtr data, size_t iter, size_t spatialDim);
+
+protected:
+  size_t channels_;
+  std::unique_ptr<Weight> scale_;
+  MatrixPtr scaleDiff_;
+  MatrixPtr normBuffer_;
+  MatrixPtr dataBuffer_;
+  MatrixPtr channelBuffer_;
+  MatrixPtr spatialBuffer_;
+  MatrixPtr sampleBuffer_;
+};
+
 }  // namespace paddle

paddle/gserver/layers/PriorBox.cpp

@@ -45,27 +45,32 @@ protected:
   MatrixPtr buffer_;
 };
 
+REGISTER_LAYER(priorbox, PriorBoxLayer);
+
 bool PriorBoxLayer::init(const LayerMap& layerMap,
                          const ParameterMap& parameterMap) {
   Layer::init(layerMap, parameterMap);
   auto pbConf = config_.inputs(0).priorbox_conf();
+  std::vector<real> tmp;
+  aspectRatio_.push_back(1.);
   std::copy(pbConf.min_size().begin(),
             pbConf.min_size().end(),
             std::back_inserter(minSize_));
   std::copy(pbConf.max_size().begin(),
             pbConf.max_size().end(),
             std::back_inserter(maxSize_));
-  std::copy(pbConf.aspect_ratio().begin(),
-            pbConf.aspect_ratio().end(),
-            std::back_inserter(aspectRatio_));
   std::copy(pbConf.variance().begin(),
             pbConf.variance().end(),
             std::back_inserter(variance_));
+  std::copy(pbConf.aspect_ratio().begin(),
+            pbConf.aspect_ratio().end(),
+            std::back_inserter(tmp));
   // flip
-  int inputRatioLength = aspectRatio_.size();
-  for (int index = 0; index < inputRatioLength; index++)
-    aspectRatio_.push_back(1 / aspectRatio_[index]);
-  aspectRatio_.push_back(1.);
+  int inputRatioLength = tmp.size();
+  for (int index = 0; index < inputRatioLength; index++) {
+    aspectRatio_.push_back(tmp[index]);
+    aspectRatio_.push_back(1 / tmp[index]);
+  }
   numPriors_ = aspectRatio_.size();
   if (maxSize_.size() > 0) numPriors_++;
   return true;
@@ -94,12 +99,12 @@ void PriorBoxLayer::forward(PassType passType) {
     for (int w = 0; w < layerWidth; ++w) {
       real centerX = (w + 0.5) * stepW;
       real centerY = (h + 0.5) * stepH;
-      int minSize = 0;
+      real minSize = 0;
       for (size_t s = 0; s < minSize_.size(); s++) {
         // first prior.
         minSize = minSize_[s];
-        int boxWidth = minSize;
-        int boxHeight = minSize;
+        real boxWidth = minSize;
+        real boxHeight = minSize;
         // xmin, ymin, xmax, ymax.
         tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth;
         tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight;
@@ -112,7 +117,7 @@ void PriorBoxLayer::forward(PassType passType) {
           CHECK_EQ(minSize_.size(), maxSize_.size());
           // second prior.
           for (size_t s = 0; s < maxSize_.size(); s++) {
-            int maxSize = maxSize_[s];
+            real maxSize = maxSize_[s];
             boxWidth = boxHeight = sqrt(minSize * maxSize);
             tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth;
             tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight;
@@ -145,6 +150,5 @@ void PriorBoxLayer::forward(PassType passType) {
   MatrixPtr outV = getOutputValue();
   outV->copyFrom(buffer_->data_, dim * 2);
 }
-REGISTER_LAYER(priorbox, PriorBoxLayer);
 
 }  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1642,6 +1642,25 @@ TEST(Layer, PadLayer) {
   }
 }
 
+TEST(Layer, CrossChannelNormLayer) {
+  TestConfig config;
+  config.layerConfig.set_type("norm");
+  config.layerConfig.set_size(100);
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  NormConfig* norm = input->mutable_norm_conf();
+  norm->set_norm_type("cross-channel-norm");
+  norm->set_channels(10);
+  norm->set_size(100);
+  norm->set_scale(0);
+  norm->set_pow(0);
+  norm->set_blocked(0);
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 100, 10});
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "cross-channel-norm", 10, false, useGpu, false, 5);
+  }
+}
+
 TEST(Layer, smooth_l1) {
   TestConfig config;
   config.layerConfig.set_type("smooth_l1");

paddle/math/BaseMatrix.cu

@@ -1453,6 +1453,24 @@ void BaseMatrixT<T>::divRowVector(BaseMatrixT& b) {
               true_type() /* bAsRowVector */, false_type());
 }
 
+template<class T>
+void BaseMatrixT<T>::mulColVector(BaseMatrixT& b) {
+  MatrixOffset offset(0, 0, 0, 0);
+  int numRows = height_;
+  int numCols = width_;
+  applyBinary(binary::DotMul<T>(), b, numRows, numCols, offset,
+              false_type(), true_type() /* bAsColVector */);
+}
+
+template<class T>
+void BaseMatrixT<T>::divColVector(BaseMatrixT& b) {
+  MatrixOffset offset(0, 0, 0, 0);
+  int numRows = height_;
+  int numCols = width_;
+  applyBinary(binary::DotDiv<T>(), b, numRows, numCols, offset,
+              false_type(), true_type() /* bAsColVector */);
+}
+
 template<>
 template <class Agg>
 int BaseMatrixT<real>::applyRow(Agg agg, BaseMatrixT& b) {

paddle/math/BaseMatrix.h

@@ -545,6 +545,9 @@ public:
   void mulRowVector(BaseMatrixT& b);
   void divRowVector(BaseMatrixT& b);
 
+  void mulColVector(BaseMatrixT& b);
+  void divColVector(BaseMatrixT& b);
+
   void addP2P(BaseMatrixT& b);
 
   /**

paddle/math/tests/test_BaseMatrix.cpp

@@ -110,6 +110,8 @@ TEST(BaseMatrix, BaseMatrix) {
       compare(&BaseMatrix::addRowVector);
       compare(&BaseMatrix::mulRowVector);
       compare(&BaseMatrix::divRowVector);
+      compare(&BaseMatrix::mulColVector);
+      compare(&BaseMatrix::divColVector);
       compare(&BaseMatrix::addP2P);
       compare(&BaseMatrix::invSqrt);
     }

python/paddle/trainer/config_parser.py

@@ -1220,9 +1220,11 @@ def parse_image(image, input_layer_name, image_conf):
 
 def parse_norm(norm, input_layer_name, norm_conf):
     norm_conf.norm_type = norm.norm_type
-    config_assert(norm.norm_type in ['rnorm', 'cmrnorm-projection'],
-                  "norm-type %s is not in [rnorm, 'cmrnorm-projection']" %
-                  norm.norm_type)
+    config_assert(
+        norm.norm_type in
+        ['rnorm', 'cmrnorm-projection', 'cross-channel-norm'],
+        "norm-type %s is not in [rnorm, cmrnorm-projection, cross-channel-norm]"
+        % norm.norm_type)
     norm_conf.channels = norm.channels
     norm_conf.size = norm.size
     norm_conf.scale = norm.scale
@@ -1898,6 +1900,9 @@ class NormLayer(LayerBase):
                        norm_conf)
             self.set_cnn_layer(name, norm_conf.output_y, norm_conf.output_x,
                                norm_conf.channels, False)
+            if norm_conf.norm_type == "cross-channel-norm":
+                self.create_input_parameter(0, norm_conf.channels,
+                                            [norm_conf.channels, 1])
 
 
 @config_layer('pool')

python/paddle/trainer_config_helpers/layers.py

@@ -112,6 +112,7 @@ __all__ = [
     'out_prod_layer',
     'print_layer',
     'priorbox_layer',
+    'cross_channel_norm_layer',
     'spp_layer',
     'pad_layer',
     'eos_layer',
@@ -1008,6 +1009,46 @@ def priorbox_layer(input,
         size=size)
 
 
+@wrap_name_default("cross_channel_norm")
+def cross_channel_norm_layer(input, name=None, param_attr=None):
+    """
+    Normalize a layer's output. This layer is necessary for ssd.
+    This layer applys normalize across the channels of each sample to
+    a conv layer's output and scale the output by a group of trainable
+    factors which dimensions equal to the channel's number.
+
+    :param name: The Layer Name.
+    :type name: basestring
+    :param input: The input layer.
+    :type input: LayerOutput
+    :param param_attr: The Parameter Attribute|list.
+    :type param_attr: ParameterAttribute
+    :return: LayerOutput
+    """
+    assert input.num_filters is not None
+    Layer(
+        name=name,
+        type=LayerType.NORM_LAYER,
+        inputs=[
+            Input(
+                input.name,
+                norm=Norm(
+                    norm_type="cross-channel-norm",
+                    channels=input.num_filters,
+                    size=input.size,
+                    scale=0,
+                    pow=0,
+                    blocked=0),
+                **param_attr.attr)
+        ])
+    return LayerOutput(
+        name,
+        LayerType.NORM_LAYER,
+        parents=input,
+        num_filters=input.num_filters,
+        size=input.size)
+
+
 @wrap_name_default("seq_pooling")
 @wrap_bias_attr_default(has_bias=False)
 @wrap_param_default(['pooling_type'], default_factory=lambda _: MaxPooling())