add deformable psroi pooling by cjt222 · Pull Request #17827 · PaddlePaddle/Paddle

cjt222 · 2019-06-04T07:30:15Z

1、性能测试

测试环境：系统为ubuntu16.04，显卡为v100
(1) Deformable PSROIPooling 和PSROI Pool对比
输入：
1）input为随机生成的维度为[5, 128, 64, 64]的数据
2）rois为随机生成的16个框，维度为[16, 4]，LOD分布为[4,3,2,2,5]
3）trans 为固定维度的[16, 2, 8, 8]，值全部为0.5
4）参数设置：
> no_trans=0, spatial_scale=1.0, output_channels=2,
group_size=[1,1], pooled_height=8, pooled_width=8, part_size=[4, 4],
sample_per_part=4, trans_std=0.1

Event	Call	Total	CPU Time (Ratio)	GPU Time (Ratio)	Min	Max	Ave
psroi_pool	100	7.36984	6.778908 (0.919818)	0.590927 (0.080182)	0.067989	0.126735	0.0736984
psroi_pool_grad	100	9.12887	7.187817 (0.787372)	1.941052 (0.212628)	0.084644	0.104966	0.0912887
deformable_psroi_pooling	100	10.3871	8.548373 (0.822981)	1.838717 (0.177019)	0.09416	0.131347	0.103871
deformable_psroi_pooling_grad	100	13.5551	8.890796 (0.655902)	4.664282 (0.344098)	0.126387	0.159987	0.135551

(2) Deformable PSROIPooling和PSROI Pool对比
输入：
1）input为随机生成的维度为[5, 128, 64, 64]的数据
2）rois为随机生成的16个框，维度为[16, 4]，LOD分布为[4,3,2,2,5]
3）trans 为固定维度的[16, 2, 8, 8]，值全部为0.5
4）参数设置：
>no_trans=1, spatial_scale=1.0, output_channels=128,
group_size=[1,1], pooled_height=8, pooled_width=8, part_size=[4, 4],
sample_per_part=4, trans_std=0.1

Event	Call	Total	CPU Time (Ratio)	GPU Time (Ratio)	Min	Max	Ave
roi_pool	100	9.0469	7.791581 (0.861243)	1.255322 (0.138757)	0.07903	0.151427	0.090469
roi_pool_grad	100	10.018	7.788079 (0.777406)	2.229958 (0.222594)	0.086172	0.177273	0.10018
deformable_psroi_pooling	100	11.9029	8.883906 (0.746367)	3.018958 (0.25253633)	0.105524	0.148849	0.119029
deformable_psroi_pooling_grad	100	21.2068	9.212682 (0.434421)	11.994113 (0.565579)	0.198663	0.268479	0.212068

2、精度测试
请进入http://agroup.baidu.com/share/md/0f274789bcde46118202deb1fc5b6343 查看
3、API预览文档

heavengate · 2019-06-04T08:48:45Z

paddle/fluid/operators/deformable_psroi_pooling_op.cc

+    AddInput("Input",
+             "(Tensor), "
+             "the input of Deformable PSROIPooling. "
+             "The format of input tensor is NCHW. Where N is batch size, "


The shape of input tensor is [N, C, H, W]

heavengate · 2019-06-04T08:54:15Z

paddle/fluid/operators/deformable_psroi_pooling_op.cu

+       i += blockDim.x * gridDim.x)
+
+const int CUDA_NUM_THREADS = 1024;
+inline int GET_BLOCKS(const int N) {


static inline int

heavengate · 2019-06-04T08:57:28Z

paddle/fluid/operators/deformable_psroi_pooling_op.cu

+  int x2 = ceil(x);
+  int y1 = floor(y);
+  int y2 = ceil(y);
+  T dist_x = (T)(x - x1);


static_cast, 下同

heavengate · 2019-06-04T09:00:01Z

paddle/fluid/operators/deformable_psroi_pooling_op.cu

+    const int group_width, const int part_height, const int part_width,
+    const int num_classes, const int channels_each_class, T* top_data,
+    T* top_count, int* roi_batch_id_data) {
+  CUDA_KERNEL_LOOP(index, count) {


please seperate following code with blank lines

每次改完这行，提交时会被pre-commit重新修正回来，貌似是clang-format的原因

嗯，是说函数里的代码用空行分一下段，现在连在一起代码看上去会不清晰，下面的函数里也是

已在代码间添加空行区分模块

heavengate · 2019-06-04T09:01:16Z

paddle/fluid/operators/deformable_psroi_pooling_op.cu

+    Tensor* top_count = ctx.Output<Tensor>("TopCount");
+    top_count->mutable_data<T>(ctx.GetPlace());
+    PADDLE_ENFORCE_EQ(top_count->dims(), out->dims(),
+                      "number of rois should be same with number of output");


better to check this in InferShape

heavengate · 2019-06-04T09:10:37Z

python/paddle/fluid/layers/nn.py

+
+    Examples:
+
+            input = fluid.layers.data(name="input",


heavengate · 2019-06-04T09:11:53Z

python/paddle/fluid/tests/unittests/test_deformable_psroi_pooling.py

+from op_test import OpTest
+
+
+class TestDeformablePSROIPoolOp(OpTest):


请针对几个attr取不同值多加几个测试用例

已添加了三组样例

heavengate · 2019-06-04T09:12:20Z

paddle/fluid/operators/deformable_psroi_pooling_op.cc

+    AddAttr<int>("no_trans",
+                 "(int), "
+                 "whether add offset to get new value or not while roi "
+                 "pooling, which value is 0 or 1");


如果有默认值，最好加一下

默认值为0，已添加

modify format

… add_deformable_psroi_pooling

add API.spec

heavengate

Please add python api test in python/paddle/fluid/tests/unittests/test_layers.py

run ci again

cjt222 · 2019-06-05T04:36:25Z

test_layers.py已添加

update API.spec

xsrobin

LGTM

qingqing01 · 2019-06-10T06:12:10Z

python/paddle/fluid/layers/nn.py

+       trans (Variable): Offset of features on ROIs while pooling.The format is NCHW, where 
+                         N is number of ROIs, C is number of channels, which indicate the offset distance 
+                         in the x and y directions, H is pooled height, and W is pooled width.
+       no_trans(integer): Whether add offset to get new value or not while roi pooling, which 


why not bool ?

和前面统一，(前面留一个空格。

已改成bool变量

qingqing01 · 2019-06-10T06:15:02Z

python/paddle/fluid/layers/nn.py

+                             trans,
+                             no_trans=0,
+                             spatial_scale=1.0,
+                             output_channels=64,


why the default output_channels is 64.

change default value to input channels

qingqing01 · 2019-06-10T06:15:41Z

python/paddle/fluid/layers/nn.py

+       output_channels(integer): The number of output channels, which should be less than input channels.
+                                 Deformable roi pooling requires output_channels = input_channels, while 
+                                 deformable psroi pooling requires output_channels = input_channels *
+                                 pooled_height * pooled_width. Default: 64.


Why default is 64?

qingqing01 · 2019-06-10T06:47:18Z

python/paddle/fluid/layers/nn.py

+                                 Deformable roi pooling requires output_channels = input_channels, while 
+                                 deformable psroi pooling requires output_channels = input_channels *
+                                 pooled_height * pooled_width. Default: 64.
+       group_size(list): The number of groups which input channels are divided.(eg.number of input channels 


qingqing01 · 2019-06-10T06:49:36Z

python/paddle/fluid/layers/nn.py

+                                                     output_channels=3,
+                                                     group_size=(1, 1),
+                                                     pooled_height=8,
+                                                     pooled_width=8,


上面描述：
output_channels = input_channels * pooled_height * pooled_width.

这里例子的设置不是这个关系。

另外，output_channels， pooled_height， pooled_width这3个要同时设置吗？可以推理得到output_channels吧。

这个应该需要看用户需求，三个量需要用户同时设置。如果用户想用deformable roi pooling这个功能的话，推理得到就实现不了这个功能了。

这个例子写的是deformable roi pool的例子，输入和输出的channels一致。

已将例子改为deformable psroi pooling的形式，另外修正了input_channels = output_channels * pooled_height * pooled_width.的关系。

modify comment

modift comment

update API.spec

modify comment

add inference in nn.py

update API.spec

qingqing01

LG for API.

But I see some duplicate code and some irregular code in C++ code.

resolve confict

qingqing01

LG for API

update API.spec

xsrobin

LGTM

cjt222 added 3 commits June 4, 2019 07:24

add deformable psroi pooling

122fa25

test=develop

5e0b79f

test=develop

784fd00

cjt222 mentioned this pull request Jun 4, 2019

Deformable PSROI Pooling OP #17827 #17669

Closed

qingqing01 requested review from heavengate, jiangjiajun and qingqing01 June 4, 2019 08:46

heavengate reviewed Jun 4, 2019

View reviewed changes

cjt222 added 5 commits June 4, 2019 12:08

test=develop

39baf4c

modify format

fix bug

e0c4978

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

4bfb3f7

… add_deformable_psroi_pooling

test=develop run ci

8ab37cf

test=develop

78fdf8b

add API.spec

heavengate reviewed Jun 4, 2019

View reviewed changes

cjt222 added 7 commits June 4, 2019 13:51

add test_layers.py

e540ccd

run ci again

8951393

test=develop

cb32eda

run ci again

run ci again

74cfa56

test=develop

ac74a0f

run ci again

test=develop

f0a2529

run ci again

test=develop

c533c0a

run ci again

add space between two lines

0394d15

test=develop

5e69a67

update API.spec

heavengate previously approved these changes Jun 10, 2019

View reviewed changes

cjt222 closed this Jun 10, 2019

cjt222 reopened this Jun 10, 2019

xsrobin previously approved these changes Jun 10, 2019

View reviewed changes

qingqing01 reviewed Jun 10, 2019

View reviewed changes

test=develop

bea535b

modify comment

cjt222 dismissed stale reviews from xsrobin and heavengate via bea535b June 10, 2019 07:57

cjt222 added 4 commits June 10, 2019 08:00

test=develop

5a2ec63

modift comment

test=develop

3b757e8

modift comment

test=develop

9c1072e

update API.spec

test=develop

dc316b0

modify comment

cjt222 closed this Jun 10, 2019

cjt222 reopened this Jun 10, 2019

cjt222 added 2 commits June 10, 2019 10:57

test=develop

02e708d

add inference in nn.py

test=develop

dd33ccc

update API.spec

qingqing01 reviewed Jun 10, 2019

View reviewed changes

cjt222 added 2 commits June 11, 2019 08:06

test=develop

6f08eef

resolve confict

test=develop

b5a1c98

resolve confict

qingqing01 previously approved these changes Jun 11, 2019

View reviewed changes

test=develop

4ddd744

update API.spec

cjt222 dismissed qingqing01’s stale review via 4ddd744 June 11, 2019 09:33

qingqing01 approved these changes Jun 11, 2019

View reviewed changes

xsrobin approved these changes Jun 11, 2019

View reviewed changes

cjt222 merged commit 871af28 into PaddlePaddle:develop Jun 11, 2019

cjt222 deleted the add_deformable_psroi_pooling branch June 11, 2019 20:56

This was referenced Jun 12, 2019

add deformable roi pooling (#17827) #18012

Closed

Cherry pick add deformable roi pooling #18014

Closed

cherry pick for deform roi pooling. #18134

Merged

		from op_test import OpTest


		class TestDeformablePSROIPoolOp(OpTest):

Conversation

cjt222 commented Jun 4, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

1、性能测试

Uh oh!

heavengate Jun 4, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

heavengate Jun 4, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

cjt222 Jun 5, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

heavengate left a comment

Choose a reason for hiding this comment

Uh oh!

cjt222 commented Jun 5, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

xsrobin left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

cjt222 Jun 10, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

cjt222 Jun 10, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

cjt222 commented Jun 4, 2019 •

edited

Loading

heavengate Jun 4, 2019 •

edited

Loading

heavengate Jun 4, 2019 •

edited

Loading

cjt222 Jun 5, 2019 •

edited

Loading

cjt222 commented Jun 5, 2019 •

edited

Loading

cjt222 Jun 10, 2019 •

edited

Loading

cjt222 Jun 10, 2019 •

edited

Loading