add ROIPooling for Fast(er) R-CNN

paddle/gserver/layers/ROIPoolLayer.cpp

@@ -0,0 +1,154 @@
+/* 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 "ROIPoolLayer.h"
+
+namespace paddle {
+
+REGISTER_LAYER(roi_pool, ROIPoolLayer);
+
+bool ROIPoolLayer::init(const LayerMap& layerMap,
+                        const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+
+  const ROIPoolConfig& layerConf = config_.inputs(0).roi_pool_conf();
+  pooledWidth_ = layerConf.pooled_width();
+  pooledHeight_ = layerConf.pooled_height();
+  spatialScale_ = layerConf.spatial_scale();
+
+  return true;
+}
+
+void ROIPoolLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  const ROIPoolConfig& layerConf = config_.inputs(0).roi_pool_conf();
+  height_ = getInput(0).getFrameHeight();
+  if (!height_) height_ = layerConf.height();
+  width_ = getInput(0).getFrameWidth();
+  if (!width_) width_ = layerConf.width();
+  channels_ = getInputValue(0)->getWidth() / width_ / height_;
+
+  size_t batchSize = getInput(0).getBatchSize();
+  size_t numROIs = getInput(1).getBatchSize();
+
+  real* bottomData = getInputValue(0)->getData();
+  size_t batchOffset = getInputValue(0)->getWidth();
+  size_t channelOffset = height_ * width_;
+  real* bottomROIs = getInputValue(1)->getData();
+  size_t roiOffset = getInputValue(1)->getWidth();
+  size_t poolChannelOffset = pooledHeight_ * pooledWidth_;
+
+  resetOutput(numROIs, channels_ * pooledHeight_ * pooledWidth_);
+  real* outputData = getOutputValue()->getData();
+  Matrix::resizeOrCreate(maxIdxs_,
+                         numROIs,
+                         channels_ * pooledHeight_ * pooledWidth_,
+                         false,
+                         false);
+  real* argmaxData = maxIdxs_->getData();
+
+  size_t uZero = 0;
+  size_t uOne = 1;
+
+  for (size_t n = 0; n < numROIs; ++n) {
+    size_t roiBatchIdx = bottomROIs[0];
+    size_t roiStartW = std::round(bottomROIs[1] * spatialScale_);
+    size_t roiStartH = std::round(bottomROIs[2] * spatialScale_);
+    size_t roiEndW = std::round(bottomROIs[3] * spatialScale_);
+    size_t roiEndH = std::round(bottomROIs[4] * spatialScale_);
+    CHECK_GE(roiBatchIdx, 0);
+    CHECK_LT(roiBatchIdx, batchSize);
+    size_t roiHeight = std::max(roiEndH - roiStartH + 1, uOne);
+    size_t roiWidth = std::max(roiEndW - roiStartW + 1, uOne);
+    real binSizeH =
+        static_cast<real>(roiHeight) / static_cast<real>(pooledHeight_);
+    real binSizeW =
+        static_cast<real>(roiWidth) / static_cast<real>(pooledWidth_);
+    real* batchData = bottomData + batchOffset * roiBatchIdx;
+    for (size_t c = 0; c < channels_; ++c) {
+      for (size_t ph = 0; ph < pooledHeight_; ++ph) {
+        for (size_t pw = 0; pw < pooledWidth_; ++pw) {
+          size_t hstart = static_cast<size_t>(std::floor(ph * binSizeH));
+          size_t wstart = static_cast<size_t>(std::floor(pw * binSizeW));
+          size_t hend = static_cast<size_t>(std::ceil((ph + 1) * binSizeH));
+          size_t wend = static_cast<size_t>(std::ceil((pw + 1) * binSizeW));
+          hstart = std::min(std::max(hstart + roiStartH, uZero), height_);
+          wstart = std::min(std::max(wstart + roiStartW, uZero), width_);
+          hend = std::min(std::max(hend + roiStartH, uZero), height_);
+          wend = std::min(std::max(wend + roiStartW, uZero), width_);
+
+          bool isEmpty = (hend <= hstart) || (wend <= wstart);
+          size_t poolIndex = ph * pooledWidth_ + pw;
+          if (isEmpty) {
+            outputData[poolIndex] = 0;
+            argmaxData[poolIndex] = -1;
+          }
+
+          for (size_t h = hstart; h < hend; ++h) {
+            for (size_t w = wstart; w < wend; ++w) {
+              size_t index = h * width_ + w;
+              if (batchData[index] > outputData[poolIndex]) {
+                outputData[poolIndex] = batchData[index];
+                argmaxData[poolIndex] = index;
+              }
+            }
+          }
+        }
+      }
+      batchData += channelOffset;
+      outputData += poolChannelOffset;
+      argmaxData += poolChannelOffset;
+    }
+    bottomROIs += roiOffset;
+  }
+}
+
+void ROIPoolLayer::backward(const UpdateCallback& callback) {
+  real* bottomROIs = getInputValue(1)->getData();
+  size_t numROIs = getInput(1).getBatchSize();
+  size_t roiOffset = getInputValue(1)->getWidth();
+
+  MatrixPtr inGrad = getInputGrad(0);
+  real* inDiffData = inGrad->getData();
+  size_t batchOffset = getInputValue(0)->getWidth();
+  size_t channelOffset = height_ * width_;
+
+  MatrixPtr outGrad = getOutputGrad();
+  real* outDiffData = outGrad->getData();
+  size_t poolChannelOffset = pooledHeight_ * pooledWidth_;
+  real* argmaxData = maxIdxs_->getData();
+
+  for (size_t n = 0; n < numROIs; ++n) {
+    size_t roiBatchIdx = bottomROIs[0];
+    real* batchDiffData = inDiffData + batchOffset * roiBatchIdx;
+    for (size_t c = 0; c < channels_; ++c) {
+      for (size_t ph = 0; ph < pooledHeight_; ++ph) {
+        for (size_t pw = 0; pw < pooledWidth_; ++pw) {
+          size_t poolIndex = ph * pooledWidth_ + pw;
+          if (argmaxData[poolIndex] > 0) {
+            size_t index = static_cast<size_t>(argmaxData[poolIndex]);
+            batchDiffData[index] += outDiffData[poolIndex];
+          }
+        }
+      }
+      batchDiffData += channelOffset;
+      outDiffData += poolChannelOffset;
+      argmaxData += poolChannelOffset;
+    }
+    bottomROIs += roiOffset;
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/ROIPoolLayer.h

@@ -0,0 +1,53 @@
+/* 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 {
+
+/**
+ * A layer used by Fast R-CNN to extract feature maps of ROIs from the last
+ * feature map.
+ * - Input: This layer needs two input layers: The first input layer is a
+ * convolution layer; The second input layer contains the ROI data which is the
+ * output of ProposalLayer in Faster R-CNN. layers for generating bbox
+ * location offset and the classification confidence. - Output: The
+ * ROIs' feature map. Reference: Shaoqing Ren, Kaiming He, Ross Girshick, and
+ * Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region
+ * Proposal
+ */
+
+class ROIPoolLayer : public Layer {
+protected:
+  size_t channels_;
+  size_t width_;
+  size_t height_;
+  size_t pooledWidth_;
+  size_t pooledHeight_;
+  real spatialScale_;
+
+  MatrixPtr maxIdxs_;
+
+public:
+  explicit ROIPoolLayer(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;
+};
+}  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1830,6 +1830,40 @@ TEST(Layer, CropLayer) {
   }
 }
 
+TEST(Layer, roi_pool) {
+  TestConfig config;
+  config.layerConfig.set_type("roi_pool");
+  config.biasSize = 0;
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ROIPoolConfig* roiPoolConf = input->mutable_roi_pool_conf();
+  roiPoolConf->set_pooled_width(7);
+  roiPoolConf->set_pooled_height(7);
+  roiPoolConf->set_spatial_scale(1. / 16);
+  roiPoolConf->set_width(14);
+  roiPoolConf->set_height(14);
+
+  MatrixPtr roiValue = Matrix::create(10, 10, false, false);
+  roiValue->zeroMem();
+  real* roiData = roiValue->getData();
+  for (size_t i = 0; i < roiValue->getElementCnt() / 5; ++i) {
+    *roiData++ = std::rand() % 2;
+    *roiData++ = std::rand() % 224;
+    *roiData++ = std::rand() % 224;
+    size_t xMin = static_cast<size_t>(*(roiData - 2));
+    size_t yMin = static_cast<size_t>(*(roiData - 1));
+    *roiData++ = xMin + std::rand() % (224 - xMin);
+    *roiData++ = yMin + std::rand() % (224 - yMin);
+  }
+
+  config.inputDefs.push_back({INPUT_DATA, "input", 3 * 14 * 14, {}});
+  config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "rois", roiValue, {}});
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "roi_pool", 5, false, useGpu, false);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

proto/ModelConfig.proto

@@ -289,6 +289,14 @@ message DetectionOutputConfig {
   optional uint32 width = 9 [default = 1];
 }
 
+message ROIPoolConfig {
+  required uint32 pooled_width = 1;
+  required uint32 pooled_height = 2;
+  required float spatial_scale = 3;
+  optional uint32 height = 4 [default = 1];
+  optional uint32 width = 5 [default = 1];
+}
+
 message LayerInputConfig {
   required string input_layer_name = 1;
   optional string input_parameter_name = 2;
@@ -309,6 +317,7 @@ message LayerInputConfig {
   optional RowConvConfig row_conv_conf = 15;
   optional MultiBoxLossConfig multibox_loss_conf = 16;
   optional DetectionOutputConfig detection_output_conf = 17;
+  optional ROIPoolConfig roi_pool_conf = 18;
 }
 
 message LayerConfig {

python/paddle/trainer/config_parser.py

@@ -1732,6 +1732,17 @@ class DetectionOutputLayer(LayerBase):
         self.config.size = size
 
 
+@config_layer('roi_pool')
+class ROIPoolLayer(LayerBase):
+    def __init__(self, name, inputs, pooled_width, pooled_height,
+                 spatial_scale):
+        super(ROIPoolLayer, self).__init__(name, 'roi_pool', 0, inputs)
+        config_assert(len(inputs) == 2, 'ROIPoolLayer must have 2 inputs')
+        self.config.inputs[0].roi_pool_conf.pooled_width = pooled_width
+        self.config.inputs[0].roi_pool_conf.pooled_height = pooled_height
+        self.config.inputs[0].roi_pool_conf.spatial_scale = spatial_scale
+
+
 @config_layer('data')
 class DataLayer(LayerBase):
     def __init__(self, name, size, height=None, width=None, device=None):

python/paddle/trainer_config_helpers/layers.py

@@ -117,6 +117,7 @@ __all__ = [
     'cross_channel_norm_layer',
     'multibox_loss_layer',
     'detection_output_layer',
+    'roi_pool_layer',
     'spp_layer',
     'pad_layer',
     'eos_layer',
@@ -201,6 +202,7 @@ class LayerType(object):
     PRIORBOX_LAYER = 'priorbox'
     MULTIBOX_LOSS_LAYER = 'multibox_loss'
     DETECTION_OUTPUT_LAYER = 'detection_output'
+    ROI_POOL_LAYER = 'roi_pool'
 
     CTC_LAYER = 'ctc'
     WARP_CTC_LAYER = 'warp_ctc'
@@ -1200,6 +1202,41 @@ def detection_output_layer(input_loc,
         name, LayerType.DETECTION_OUTPUT_LAYER, parents=parents, size=size)
 
 
+@wrap_name_default("roi_pool")
+def roi_pool_layer(input,
+                   rois,
+                   pooled_width,
+                   pooled_height,
+                   spatial_scale,
+                   name=None):
+    """
+    A layer used by Fast R-CNN to extract feature maps of ROIs from the last
+    feature map.
+
+    :param name: The Layer Name.
+    :type name: basestring
+    :param input: The input layer.
+    :type input: LayerOutput.
+    :param rois: The input ROIs' data.
+    :type rois: LayerOutput.
+    :param pooled_width: The width after pooling.
+    :type pooled_width: int
+    :param pooled_height: The height after pooling.
+    :type pooled_height: int
+    :param spatial_scale: The spatial scale between the image and feature map.
+    :type spatial_scale: float
+    :return: LayerOutput
+    """
+    Layer(
+        name=name,
+        type=LayerType.ROI_POOL_LAYER,
+        inputs=[input.name, rois.name],
+        pooled_width=pooled_width,
+        pooled_height=pooled_height,
+        spatial_scale=spatial_scale)
+    return LayerOutput(name, LayerType.ROI_POOL_LAYER, parents=[input, rois])
+
+
 @wrap_name_default("cross_channel_norm")
 def cross_channel_norm_layer(input, name=None, param_attr=None):
     """