NormalizeLayer for SSD

paddle/gserver/layers/NormalizeLayer.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 "paddle/math/BaseMatrix.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+/**
+ * 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.
+ * - Input: One and only one input layer are accepted. The input layer must be
+ *        be a data output layer.
+ * - 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 NormalizeLayer : public Layer {
+public:
+  explicit NormalizeLayer(const LayerConfig& config) : Layer(config) {}
+  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+
+  void forward(PassType passType);
+  void backward(const UpdateCallback& callback);
+
+protected:
+  size_t channels_;
+  std::unique_ptr<Weight> scale_;
+  MatrixPtr scaleDiff_;
+  MatrixPtr normBuffer_;
+  MatrixPtr dataBuffer_;
+  MatrixPtr channelBuffer_;
+  MatrixPtr spatialBuffer_;
+  MatrixPtr sampleBuffer_;
+};
+
+REGISTER_LAYER(normalize, NormalizeLayer);
+
+bool NormalizeLayer::init(const LayerMap& layerMap,
+                          const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+  CHECK(parameters_[0]);
+  channels_ = config_.num_filters();
+  scale_.reset(new Weight(channels_, 1, parameters_[0]));
+  return true;
+}
+
+void NormalizeLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  auto in = getInput(0);
+  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(channelBuffer_, channels_, 1, false, useGpu_);
+  Matrix::resizeOrCreate(sampleBuffer_, channels_, spatialDim, false, useGpu_);
+  Matrix::resizeOrCreate(normBuffer_, batchSize, spatialDim, false, useGpu_);
+  normBuffer_->zeroMem();
+  spatialBuffer_->zeroMem();
+  sampleBuffer_->zeroMem();
+  dataBuffer_->zeroMem();
+  // add eps to avoid overflow
+  normBuffer_->addScalar(*normBuffer_, 1e-6);
+  channelBuffer_->resetOne();
+  inV->square2(*dataBuffer_);
+  for (size_t i = 0; i < batchSize; i++) {
+    spatialBuffer_->zeroMem();
+    MatrixPtr inTmp = Matrix::create(
+        inV->getData() + i * dataDim, channels_, spatialDim, false, useGpu_);
+    MatrixPtr dataTmp = Matrix::create(dataBuffer_->getData() + i * dataDim,
+                                       channels_,
+                                       spatialDim,
+                                       false,
+                                       useGpu_);
+    MatrixPtr outTmp = Matrix::create(
+        outV->getData() + i * dataDim, channels_, spatialDim, false, useGpu_);
+    MatrixPtr normTmp = Matrix::create(
+        normBuffer_->getData() + i * spatialDim, 1, spatialDim, false, useGpu_);
+    // compute norm.
+    spatialBuffer_->sumCols(*dataTmp, 1, 1);
+    spatialBuffer_->sqrt2(*spatialBuffer_);
+    normTmp->copyFrom(*spatialBuffer_);
+    sampleBuffer_->mul(*channelBuffer_, *spatialBuffer_, 1., 0.);
+    sampleBuffer_->dotDiv(*inTmp, *sampleBuffer_);
+    outTmp->copyFrom(*sampleBuffer_);
+
+    // scale the layer.
+    spatialBuffer_->resetOne();
+    sampleBuffer_->mul(*scale_->getW(), *spatialBuffer_, 1., 0.);
+    outTmp->dotMul(*outTmp, *sampleBuffer_);
+  }
+}
+
+void NormalizeLayer::backward(const UpdateCallback& callback) {
+  MatrixPtr inG = getInputGrad(0);
+  MatrixPtr inV = getInputValue(0);
+  MatrixPtr outG = getOutputGrad();
+  MatrixPtr outV = getOutputValue();
+
+  auto in = getInput(0);
+  size_t batchSize = inG->getHeight();
+  size_t dataDim = inG->getWidth();
+  size_t spatialDim = dataDim / channels_;
+
+  bool syncFlag = hl_get_sync_flag();
+  dataBuffer_->dotMul(*outG, *outV);
+  Matrix::resizeOrCreate(scaleDiff_, channels_, 1, false, useGpu_);
+  scaleDiff_->zeroMem();
+  for (size_t i = 0; i < batchSize; i++) {
+    spatialBuffer_->zeroMem();
+    channelBuffer_->zeroMem();
+    // propagate to param.
+    MatrixPtr dataBufferTmp =
+        Matrix::create(dataBuffer_->getData() + i * dataDim,
+                       channels_,
+                       spatialDim,
+                       false,
+                       useGpu_);
+    const MatrixPtr inValueTmp = Matrix::create(
+        inV->getData() + i * dataDim, channels_, spatialDim, false, useGpu_);
+    const MatrixPtr outGradTmp = Matrix::create(
+        outG->getData() + i * dataDim, channels_, spatialDim, false, useGpu_);
+    MatrixPtr inGradTmp = Matrix::create(
+        inG->getData() + i * dataDim, channels_, spatialDim, false, useGpu_);
+    const MatrixPtr normTmp = Matrix::create(
+        normBuffer_->getData() + i * spatialDim, 1, spatialDim, false, useGpu_);
+    channelBuffer_->sumRows(*dataBufferTmp, 1, 1);
+    channelBuffer_->dotDiv(*channelBuffer_, *(scale_->getW()));
+    // store a / scale[i] in scaleDiff_ temporary
+    scaleDiff_->add(*channelBuffer_, 1.);
+
+    sampleBuffer_->dotMul(*inValueTmp, *outGradTmp);
+    spatialBuffer_->sumCols(*sampleBuffer_, 1., 1.);
+    // scale the grad
+    channelBuffer_->resetOne();
+    sampleBuffer_->mul(*channelBuffer_, *spatialBuffer_, 1., 0.);
+
+    inGradTmp->dotMul(*inValueTmp, *sampleBuffer_);
+    // divide by square of norm
+    spatialBuffer_->dotMul(*normTmp, *normTmp);
+    sampleBuffer_->mul(*channelBuffer_, *spatialBuffer_, 1., 0.);
+    inGradTmp->dotDiv(*inGradTmp, *sampleBuffer_);
+    // subtract
+    inGradTmp->add(*outGradTmp, -1, 1);
+    // divide by norm
+    sampleBuffer_->mul(*channelBuffer_, *normTmp, 1., 0.);
+    inGradTmp->dotDiv(*inGradTmp, *sampleBuffer_);
+    // scale the diff
+    spatialBuffer_->resetOne();
+    sampleBuffer_->mul(*scale_->getW(), *spatialBuffer_, 1., 0.);
+    inGradTmp->dotMul(*inGradTmp, *sampleBuffer_);
+  }
+  // updata scale
+  if (scale_->getWGrad()) scale_->getWGrad()->copyFrom(*scaleDiff_);
+  hl_set_sync_flag(false);
+  hl_set_sync_flag(syncFlag);
+  scale_->getParameterPtr()->incUpdate(callback);
+}
+
+}  // namespace paddle

paddle/gserver/layers/PriorBox.cpp

@@ -20,7 +20,7 @@ namespace paddle {
 /**
  * @brief A layer for generating priorbox locations and variances.
  * - Input: Two and only two input layer are accepted. The input layer must be
- *        be a data output layer and a convolution output layer.
+ *          be a data output layer and a convolution output layer.
  * - Output: The priorbox locations and variances of the input data.
  * Reference:
  *    Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed,
@@ -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

@@ -1623,6 +1623,20 @@ TEST(Layer, PadLayer) {
   }
 }
 
+TEST(Layer, NormalizeLayer) {
+  TestConfig config;
+  config.layerConfig.set_type("normalize");
+  config.layerConfig.set_size(100);
+  config.layerConfig.set_num_filters(10);
+
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 100, 10});
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "normalize", 10, false, useGpu, false, 5);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

python/paddle/trainer/config_parser.py

@@ -1619,6 +1619,16 @@ class PriorBoxLayer(LayerBase):
         self.config.size = size
 
 
+@config_layer('normalize')
+class NormalizeLayer(LayerBase):
+    def __init__(self, name, inputs, size, num_filters, **xargs):
+        super(NormalizeLayer, self).__init__(name, 'normalize', 0, inputs,
+                                             **xargs)
+        self.config.size = size
+        self.config.num_filters = num_filters
+        self.create_input_parameter(0, num_filters, [num_filters, 1])
+
+
 @config_layer('data')
 class DataLayer(LayerBase):
     def __init__(self, name, size, height=None, width=None, device=None):

python/paddle/trainer_config_helpers/layers.py

@@ -111,6 +111,7 @@ __all__ = [
     'out_prod_layer',
     'print_layer',
     'priorbox_layer',
+    'normalize_layer',
     'spp_layer',
     'pad_layer',
     'eos_layer',
@@ -184,6 +185,7 @@ class LayerType(object):
 
     PRINT_LAYER = "print"
     PRIORBOX_LAYER = "priorbox"
+    NORMALIZE_LAYER = "normalize"
 
     CTC_LAYER = "ctc"
     WARP_CTC_LAYER = "warp_ctc"
@@ -998,6 +1000,35 @@ def priorbox_layer(input,
         size=size)
 
 
+@wrap_name_default("normalize")
+def normalize_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
+    """
+    Layer(
+        name=name,
+        type=LayerType.NORMALIZE_LAYER,
+        inputs=[Input(input.name, **param_attr.attr)],
+        size=input.size,
+        num_filters=input.num_filters)
+    return LayerOutput(
+        name,
+        LayerType.NORMALIZE_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())