Add DetectionOutputLayer and MultiBoxLossLayer. #2497

pkuyym · 2017-06-18T09:18:15Z

resolves #2496

qingqing01 · 2017-06-27T05:31:52Z

paddle/gserver/layers/DetectionOutputLayer.cpp

+    locBuffer_ = locCpuBuffer_;
+    confBuffer_ = confCpuBuffer_;
+    priorValue = priorCpuValue_;
+  } else {


这里觉得，上面appendWithPermute直接操作locTmpBuffer_ 和confTmpBuffer_ ，在else分支里再给locBuffer_ 和confBuffer_ 赋值逻辑更清晰些：

locBuffer_ = locTmpBuffer_； confBuffer_ = confTmpBuffer_；

qingqing01 · 2017-06-27T05:36:30Z

paddle/gserver/layers/DetectionOutputLayer.cpp

+  }
+  confBuffer_->softmax(*confBuffer_);
+
+  size_t numPriors = priorValue->getElementCnt() / 8;


同上个PR，8、4、7这样的常量需要重新定义，避免直接在代码里用数字，不结合上下文，不能直接看懂。不过后续重构用Tensor更好实现些~

qingqing01 · 2017-06-27T05:48:54Z

paddle/gserver/layers/DetectionOutputLayer.h

+
+using std::vector;
+using std::map;
+using std::pair;


Do not use aliases for std::vector, std::map and std::pair in the header. https://google.github.io/styleguide/cppguide.html#Aliases

qingqing01 · 2017-06-27T05:49:24Z

paddle/gserver/layers/DetectionOutputLayer.h

+ * The detection output layer for a SSD detection task. This layer apply the
+ * Non-maximum suppression to the all predicted bounding box and keep the
+ * Top-K bounding boxes.
+ * - Input: This layer need three input layers: This first input layer


need -> needs

qingqing01 · 2017-06-27T11:12:04Z

paddle/gserver/layers/MultiBoxLossLayer.cpp

+          confProb_->getData()[count * numClasses_ + j] =
+              (confBuffer_->getData() + confOffset)[j];
+          confPredData.push_back((confBuffer_->getData() + confOffset)[j]);
+        }


for循环里的拷贝可以使用std::copy.

qingqing01 · 2017-06-27T11:37:18Z

python/paddle/trainer/config_parser.py

+@config_layer('multibox_loss')
+class MultiBoxLossLayer(LayerBase):
+    def __init__(self, name, inputs, input_num, num_classes, overlap_threshold,
+                 neg_pos_ratio, neg_overlap, background_id):


background_id): -> background_id, **xargs):

qingqing01 · 2017-06-27T11:38:02Z

python/paddle/trainer/config_parser.py

+class DetectionOutputLayer(LayerBase):
+    def __init__(self, name, inputs, size, input_num, num_classes,
+                 nms_threshold, nms_top_k, keep_top_k, confidence_threshold,
+                 background_id):


background_id): -> background_id, **xargs):

qingqing01 · 2017-06-27T11:43:20Z

python/paddle/trainer_config_helpers/layers.py

+    assert isinstance(input_loc, collections.Sequence)  # list or tuple
+    for each in input_loc:
+        assert isinstance(each, LayerOutput)
+        input_loc_num += 1


可以去掉前面input_loc_num=0，改成：

for each in input_loc: assert isinstance(each, LayerOutput) input_loc_num = len(input_loc)

input_conf_num同样~

这块没有Done.

qingqing01 · 2017-06-27T11:45:21Z

python/paddle/trainer_config_helpers/layers.py

+    assert isinstance(input_conf, collections.Sequence)  # list or tuple
+    for each in input_conf:
+        assert isinstance(each, LayerOutput)
+        input_conf_num += 1


这块也没有Done.

qingqing01 · 2017-06-27T11:45:58Z

python/paddle/trainer_config_helpers/layers.py

@@ -115,6 +115,8 @@
    'print_layer',
    'priorbox_layer',
    'cross_channel_norm_layer',
+    'multibox_loss_layer',
+    'detection_output_layer',


增加到文档里，同时增加单测~

All unit tests have been added. Which document?

doc/api/v2/config/layer.rst

qingqing01 · 2017-06-30T13:15:06Z

paddle/gserver/layers/MultiBoxLossLayer.cpp

+  real loss = locLoss_ + confLoss_;
+  MatrixPtr outV = getOutputValue();
+  std::vector<real> tmp(batchSize, loss);
+  outV->copyFrom(&tmp[0], batchSize);


每个样本的loss都一样？如果是，直接赋值给outV就可以，不用中间构造:

std::vector<real> tmp(batchSize, loss);

看下 BaseMatrix::assign是否可用。

qingqing01 · 2017-06-30T13:17:36Z

paddle/gserver/layers/MultiBoxLossLayer.cpp

+    size_t height = getInput(*getLocInputLayer(n)).getFrameHeight();
+    if (!height) height = layerConf.height();
+    size_t width = getInput(*getLocInputLayer(n)).getFrameWidth();
+    if (!width) width = layerConf.width();


layerConf.height()和layerConf.width()为了单测才加的，仅仅在这里注释下。

qingqing01 · 2017-06-30T13:23:32Z

paddle/gserver/layers/MultiBoxLossLayer.cpp

+    size_t height = getInput(*getLocInputLayer(n)).getFrameHeight();
+    if (!height) height = layerConf.height();
+    size_t width = getInput(*getLocInputLayer(n)).getFrameWidth();
+    if (!width) width = layerConf.width();


size_t width = getInput(*getLocInputLayer(n)).getFrameWidth();

这句理解，每个LocInputLayer的height/width可以不一样，实际也不一样吧? 但为了单测加的：

if (!width) width = layerConf.width();

确实每个LocInputLayer的height/width都一样。

qingqing01 · 2017-06-30T13:24:15Z

paddle/gserver/layers/MultiBoxLossLayer.h

+ * The loss is composed by the location loss and the confidence loss.
+ * The location loss is a smooth L1 loss and the confidence loss is
+ * a softmax loss.
+ * - Input: This layer need four input layers: This first input layer


need -> needs
This first input -> The first input

qingqing01 · 2017-06-30T13:25:22Z

paddle/gserver/layers/DetectionOutputLayer.h

+namespace paddle {
+
+/**
+ * The detection output layer for a SSD detection task. This layer apply the


apply -> applies

qingqing01 · 2017-06-30T13:36:59Z

python/paddle/trainer_config_helpers/layers.py

+    :param name: The Layer Name.
+    :type name: basestring
+    :param input_loc: The input predict location.
+    :type input_loc: LayerOutput


LayerOutput | List of LayerOutput.

qingqing01 · 2017-06-30T13:37:09Z

python/paddle/trainer_config_helpers/layers.py

+    :param input_loc: The input predict location.
+    :type input_loc: LayerOutput
+    :param input_conf: The input priorbox confidence.
+    :type input_conf: LayerOutput


LayerOutput | List of LayerOutput.

qingqing01 · 2017-06-30T13:37:53Z

python/paddle/trainer_config_helpers/layers.py

+    assert isinstance(input_loc, collections.Sequence)  # list or tuple
+    for each in input_loc:
+        assert isinstance(each, LayerOutput)
+        input_loc_num += 1


这块没有Done.

qingqing01 · 2017-06-30T13:38:05Z

python/paddle/trainer_config_helpers/layers.py

+    assert isinstance(input_conf, collections.Sequence)  # list or tuple
+    for each in input_conf:
+        assert isinstance(each, LayerOutput)
+        input_conf_num += 1


这块也没有Done.

qingqing01 · 2017-06-30T13:39:07Z

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

@@ -0,0 +1,25 @@
+from paddle.trainer_config_helpers import *


需要运行下单测，将protostr文件add进来。上同。

Add DetectionOutputLayer and MultiBoxLossLayer.

65969da

pkuyym requested a review from qingqing01 June 18, 2017 09:18

Set FLAGS_use_gpu in test_DetectionOutput.

b233ed1

qingqing01 requested changes Jun 27, 2017

View reviewed changes

pkuyym added 2 commits June 29, 2017 15:11

Follow comments, mainly use std::copy to simplify logic.

b5ab4b6

Add test for configuration and add doc.

3a0919b

qingqing01 reviewed Jun 30, 2017

View reviewed changes

Follow comments.

86543f7

qingqing01 approved these changes Jul 4, 2017

View reviewed changes

pkuyym merged commit 312ce8b into PaddlePaddle:develop Jul 4, 2017

pkuyym deleted the ssd_outloss branch July 4, 2017 15:20

Add DetectionOutputLayer and MultiBoxLossLayer. #2497

Add DetectionOutputLayer and MultiBoxLossLayer. #2497

Conversation

pkuyym commented Jun 18, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

qingqing01 Jun 27, 2017 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

qingqing01 Jun 27, 2017 •

edited