priorbox layer for Single Shot Multibox Detection Network #867

Noplz · 2016-12-13T13:05:53Z

The priorbox layer for Single Shot Multibox Detection Network. This layer doesn't need the backward propagation and the layer's result is independent of the value of the input image but the shape. The layer's result has been verified by the Caffe's implementation.

qingqing01 · 2016-12-14T01:57:49Z

@Noplz Please format the code.

You should install pre-commit and clang-format.

pip install pre-commit.

Then run following command in the Git root of Paddle,

pre-commit install

After the above has been done, the code will be automatically formatted when running git commit .

luotao1 · 2016-12-14T02:15:33Z

paddle/gserver/layers/PriorBox.cpp

+            std::back_inserter(aspectRatio_));
+  std::copy(config_.inputs(0).priorbox_conf().variance().begin(),
+            config_.inputs(0).priorbox_conf().variance().end(),
+            std::back_inserter(variance_));


可以先用一个临时变量来记录config_.inputs(0).priorbox_conf()，38-49行不用重复写那么多了

luotao1 · 2016-12-14T02:17:11Z

缺少单测。缺少GPU实现（如果在下一个PR补充，请说明）。

Xreki · 2016-12-14T02:21:17Z

paddle/gserver/layers/PriorBox.cpp

@@ -0,0 +1,137 @@
+/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.


Baidu, Inc需统一改成PaddlePaddle Authors

好这就去修改

qingqing01 · 2016-12-14T11:04:00Z

paddle/gserver/layers/PriorBox.cpp

+void PriorBoxLayer::forwardImp(const Argument& featureMap,
+                               const Argument& imageShape) {
+  int layer_width = featureMap.getFrameWidth();
+  int layer_height = featureMap.getFrameHeight();


有很多变量命名不符合Paddle规范。

好的..之前没注意到..

已修改，麻烦再review下。

qingqing01 · 2016-12-14T11:07:06Z

python/paddle/trainer/config_parser.py

+        self.config.inputs[0].priorbox_conf.variance.extend(variance)
+        self.config.size = size
+        input_layer0 = self.get_input_layer(0)
+        input_layer1 = self.get_input_layer(1)


input_layer0和input_layer1没用的话可以删掉。

qingqing01 · 2016-12-14T11:17:07Z

paddle/gserver/layers/PriorBox.cpp

+
+  MatrixPtr inV1 = imageShape.value;
+  int image_width = inV1->getElement(0, 0);
+  int image_height = inV1->getElement(0, 1);


第2个Input只存了height和width？每个样本都一样吗？可以从输入的data_layer的frameHeight 和 frameWidth获得吗？

对的每个样本都一样的印象中从data_layer没能get到frameHeight和frameWidth

如果每个样本都一样，data_layer现在可以设置image的height和width了，可以通过frameHeight和frameWidth获取到。

那我再去试一下

已修改为从data_layer的output获得image的height和width

qingqing01 · 2016-12-14T11:17:53Z

paddle/gserver/layers/PriorBox.cpp

+
+namespace paddle {
+
+class PriorBoxLayer : public Layer {


添加注释。

qingqing01 · 2016-12-14T11:20:02Z

paddle/gserver/layers/PriorBox.cpp

+  aspectRatio_.push_back(1.);
+  numPriors_ = aspectRatio_.size();
+  if (maxSize_.size() > 0) numPriors_++;
+  buffer_ = Matrix::create(1, 1, false, false);


下面有resizeOrCreate，这行可以删掉了

qingqing01 · 2016-12-14T11:26:48Z

paddle/gserver/layers/PriorBox.cpp

+  if (useGpu_) {
+    for (size_t i = 0; i < inputLayers_.size(); i++) {
+      tmpCpuInput_[i].resizeAndCopyFrom(
+          getInput(i), false, HPPL_STREAM_DEFAULT);


下面计算只用到了input的value，这里其实没有必要拷贝整个Argument。只需要拷贝value就可以了。因此 std::vector<Argument> tmpCpuInput_; 这个其实可以是两个Matrix。

如果传Matrix的话, getFrameWidth和getFrameHeight就获取不到了吧？到是可以在外面把这几个值都get出来只传这几个int型的width和height的值进去。其实这一层需要的只是第一个input matrix的shape，第二个input的matrix也只有两个值，都通过getElement拿到的话也无所谓数据在CPU或GPU，output也是通过copyFrom拿到的。所以感觉这里加if(useGpu_)也有些多余。

如果只获取shape的话，无论数据是在CPU还是GPU，通过Argument或者Matrix都可以获取到，所以input应该也不需要拷贝了。

嗯。。我还是改回最开始那样吧不需要任何copy

已修改为不需要任何copy兼容cpu和gpu

Noplz · 2016-12-15T08:09:59Z

现在还少一个单测，这个层感觉比较特殊还没想好怎么写单测好。

luotao1 · 2016-12-15T08:31:18Z

现在还少一个单测，这个层感觉比较特殊还没想好怎么写单测好。

是否可以仿照 https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/test_ConvTrans.cpp 自定义result结果？

Noplz · 2016-12-15T08:33:37Z

好的我去看看仿照写一个

qingqing01 · 2016-12-15T11:33:50Z

python/paddle/trainer_config_helpers/layers.py

+    """
+    # plus one for ratio 1.
+    num_filters = (len(aspect_ratio) * 2 + 1 + len(max_size)) * 4
+    size = (input.size / input.num_filters) * num_filters * 2


可以对第二个input即Image进行check，check必须为data_layer? config_parse.py里面或许可以check必须设置了height和width。

已添加check 都放在config_parser.py里面check了

Noplz · 2016-12-19T11:44:57Z

仿照着https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/test_ConvTrans.cpp 写了一个单测，麻烦review下，谢谢

Noplz · 2016-12-20T09:02:17Z

已添加gpu单测

qingqing01 · 2016-12-20T13:12:03Z

paddle/gserver/tests/test_PriorBox.cpp

+
+  // CPU case 1.
+  MatrixPtr result;
+  float resultData[] = {0.04,


额，由 #967 注意到，这里得改成 real，下同 :)

priorbox layer for ssd

0eac399

qingqing01 self-assigned this Dec 14, 2016

luotao1 reviewed Dec 14, 2016

View reviewed changes

qingqing01 assigned Xreki Dec 14, 2016

Xreki reviewed Dec 14, 2016

View reviewed changes

gaoyuan added 3 commits December 14, 2016 15:19

Format the priorbox code

39d689e

Add fake gpu support of the priorbox layer for the moment

9600932

Format the python file.

c007608

qingqing01 requested changes Dec 14, 2016

View reviewed changes

qingqing01 reviewed Dec 14, 2016

View reviewed changes

gaoyuan added 2 commits December 15, 2016 13:33

Fix variable name and add the annotation

707a9c9

Fix code format

520342e

qingqing01 reviewed Dec 15, 2016

View reviewed changes

gaoyuan added 6 commits December 15, 2016 20:25

Add input layer check

1048aee

Add unittest of the priorbox layer

9f990d9

Add header files

8d9f675

Add header file

cad325f

remove random flag

38723e7

remove gpu memory alloc

7dfe3bd

Add priorbox layer gpu unit test.

6f8f468

qingqing01 approved these changes Dec 20, 2016

View reviewed changes

qingqing01 requested changes Dec 20, 2016

View reviewed changes

wangkuiyi changed the title ~~priorbox layer for ssd~~ priorbox layer for Single Shot Multibox Detection Network Dec 20, 2016

Change type float to real.

f202929

qingqing01 approved these changes Dec 21, 2016

View reviewed changes

Change member variables from public to protected

8d24931

qingqing01 merged commit b53bdcd into PaddlePaddle:develop Dec 22, 2016

		@@ -0,0 +1,137 @@
		/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.

priorbox layer for Single Shot Multibox Detection Network #867

priorbox layer for Single Shot Multibox Detection Network #867

Conversation

Noplz commented Dec 13, 2016

qingqing01 commented Dec 14, 2016

Choose a reason for hiding this comment

Choose a reason for hiding this comment

luotao1 commented Dec 14, 2016

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 Dec 14, 2016 • edited Loading

Choose a reason for hiding this comment

Noplz Dec 14, 2016 • edited Loading

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

Noplz commented Dec 15, 2016

luotao1 commented Dec 15, 2016

Noplz commented Dec 15, 2016

qingqing01 Dec 15, 2016 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Noplz commented Dec 19, 2016 • edited Loading

Noplz commented Dec 20, 2016

Choose a reason for hiding this comment

Choose a reason for hiding this comment

qingqing01 Dec 14, 2016 •

edited

Loading

Noplz Dec 14, 2016 •

edited

Loading

qingqing01 Dec 15, 2016 •

edited

Loading

Noplz commented Dec 19, 2016 •

edited

Loading