[API2.0]Unify pooling function and add adaptive max pooling function #26483

shippingwang · 2020-08-20T06:01:23Z

PR types

New features

PR changes

APIs

Describe

This PR

add adaptive max pooling 2d 3d function
Uniform paddle pooling function, extract common function 1. _update_padding_nd which is consistent with conv function, 2. _expand_low_nd_padding to implement low dimension function by calling high dimension function(unsqueeze nd to n+1 d)
solve [API 2.0] add pool2d3d API,test=develop #26331 and [2.0API]Add adaptive_avg_pool_2/3d #26369 problem
- fix typo
- uniform name and order
- add deprecated fluid function
- add Shape field in layer/pooling.py
- add ref link in functional/pooling.py
- use x instead of input, out instead of output
uniform padding description, which is consistent with conv function.
Example:

padding (string|int|list|tuple): The padding size. Padding could be in one of the following fo rms.
1. A string in ['valid', 'same'].
2. An int, which means the feature map is zero-padded by the size of `padding` on every side.
3. A list[int] or tuple(int) whose length is 1, which means the feature map is zero-padded by the size of `padding[0]` on every side.
4. A list[int] or tuple(int) whose length is 2. It has the form [pad_before, pad_after].
5. A list or tuple of pairs of integers. It has the form [[pad_before, pad_after], [pad_before, pad_after], ...]. Note that, the batch dimension and channel dimension should be [0,0] or (0,0).
he default value is 0.

support padding both front and back in 1d function, although it is NOT consistent with PyTorch, It is same as PaddlePaddle conv 1d function and other pooling Nd functions.
separate adaptive 1/2/3d max/avg unittest

paddle-bot-old · 2020-08-20T06:02:26Z

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

… add_adaptive_max_pooling

WuHaobo

LGTM

willthefrog

double check data_format
double check use_cudnn, also document reasoning behind use_cudnn default value choice.
also for future references, it would be nice to have some preliminary benchmark on 1d pooling in PR description

python/paddle/nn/functional/pooling.py

willthefrog · 2020-08-27T03:37:05Z

python/paddle/nn/functional/pooling.py

-            'SAME' which is the padding algorithm. If pool padding size is a tuple or list,
-            it could be the following forms: `[pad_left, pad_right]`. If padding is non-zero,
-            then the input is implicitly zero-padded on both sides for padding number of points.
+            it must contain an integer.


1d is an integer

python/paddle/nn/functional/pooling.py

willthefrog · 2020-08-27T03:47:14Z

python/paddle/nn/functional/pooling.py


-    padding = update_padding1d(padding, "avg")
+    # use 2d to implenment 1d should expand padding in advance.
+    padding = _expand_low_nd_padding(padding)

    if in_dygraph_mode():
        output = core.ops.pool2d(


use_cudnn is different?
also, document the reasoning of use_cudnn default value.

It looks like a merged bug https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/nn/functional/pooling.py#L247 , someone mixed up use_cudnn and exclusive params, I should and already correct it in the next commit.

willthefrog · 2020-08-27T03:56:57Z

python/paddle/nn/layer/pooling.py

+            ceil_mode=self.ceil_mode,
+            count_include_pad=self.count_include_pad,
+            divisor_override=self.divisor,
+            data_format=self.data_format,


data_format works?

willthefrog · 2020-08-27T04:07:00Z

python/paddle/nn/functional/pooling.py

@@ -146,99 +168,69 @@ def avg_pool1d(x,
               count_include_pad=True,
               ceil_mode=False,
               name=None):


missing data format?

It supposes to add in this PR #26331, hmm I suggest refining it in the next PR

willthefrog

check weather custom pooling implementation is faster than cudnn for adaptive/global

if both faster, set use_cudnn to false for adaptive and global
if global is faster, set global_pooling to true and use_cudnn to false for adaptive pooling with output size 1

shippingwang · 2020-08-28T09:26:28Z

check weather custom pooling implementation is faster than cudnn for adaptive/global

if both faster, set use_cudnn to false for adaptive and global

if global is faster, set global_pooling to true and use_cudnn to false for adaptive pooling with output size 1

I will compare the performance of core.ops.[max/avg]_pool[2/3]d_with_index(), core.ops.pool[2/3]d (use_cudnn=Ture, use global_pooling=True), and core.ops.pool[2/3]d(use_cudnn=False, use_global_pooling=True) when the output size is 1 in adaptive [max/avg] pool[1/2/3]d function in next PR.

please feel free to review it again while CI is passing

willthefrog

need follow up for data format and cudnn, otherwise LGTM

XiaoguangHu01

LG API

XiaoguangHu01 · 2020-08-28T12:01:18Z

python/paddle/fluid/tests/unittests/test_adaptive_avg_pool1d.py

+    return int(np.ceil((index + 1) * input_size / output_size))
+
+
+def avg_pool1D_forward_naive(x,


phlrain

LGTM for core.ops

add adaptive max pooling

9691cbf

shippingwang added 13 commits August 20, 2020 17:58

add function

691ab3a

add class

2170025

fix conflict

61c9adc

fix conflict

fe383eb

fix conflict

5e22ce5

uniform all pooling

16ef588

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

0db9342

… add_adaptive_max_pooling

add layer

ea0212c

correct wrong upload

c2fddd3

refine

9f62609

add ref and Shape

fec8256

add deprecated

09d2470

refine name and add test

33e45ca

shippingwang changed the title ~~[API2.0][WIP]add adaptive max pooling~~ [API2.0][WIP]Uniform pooling function and add adaptive max pooling function Aug 25, 2020

shippingwang added 3 commits August 26, 2020 06:59

separate pooling unittest

768c514

remove temp files

5d64357

fix conflict

f8c82e8

shippingwang requested review from willthefrog, LDOUBLEV, baiyfbupt and XiaoguangHu01 August 26, 2020 07:16

shippingwang added 2 commits August 26, 2020 09:05

fix

36b70dc

fix unittest mask bug

861772e

shippingwang changed the title ~~[API2.0][WIP]Uniform pooling function and add adaptive max pooling function~~ [API2.0]Uniform pooling function and add adaptive max pooling function Aug 26, 2020

shippingwang requested a review from jzhang533 August 26, 2020 13:22

shippingwang added 3 commits August 27, 2020 02:59

fix unittest bug

ed4f80a

fix ceil mode bug

3bebc0e

move adaptive pool 1d to separate files

5c48e98

WuHaobo previously approved these changes Aug 27, 2020

View reviewed changes

willthefrog requested changes Aug 27, 2020

View reviewed changes

shippingwang changed the title ~~[API2.0]Uniform pooling function and add adaptive max pooling function~~ [API2.0]Unify pooling function and add adaptive max pooling function Aug 27, 2020

fix conflict

ab9e6a8

shippingwang dismissed WuHaobo’s stale review via ab9e6a8 August 27, 2020 13:30

shippingwang requested review from willthefrog, zhiqiu and raindrops2sea August 27, 2020 15:43

fix bug

dc6ba52

willthefrog reviewed Aug 28, 2020

View reviewed changes

fix unittest

366b82c

shippingwang requested a review from willthefrog August 28, 2020 09:27

willthefrog approved these changes Aug 28, 2020

View reviewed changes

XiaoguangHu01 approved these changes Aug 28, 2020

View reviewed changes

shippingwang requested a review from phlrain August 28, 2020 13:54

phlrain approved these changes Aug 28, 2020

View reviewed changes

raindrops2sea approved these changes Aug 29, 2020

View reviewed changes

shippingwang merged commit db68e08 into PaddlePaddle:develop Aug 29, 2020

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[API2.0]Unify pooling function and add adaptive max pooling function #26483

[API2.0]Unify pooling function and add adaptive max pooling function #26483

shippingwang commented Aug 20, 2020 •

edited

paddle-bot-old bot commented Aug 20, 2020

WuHaobo left a comment

willthefrog left a comment •

edited

willthefrog Aug 27, 2020

shippingwang Aug 27, 2020

willthefrog Aug 27, 2020

shippingwang Aug 27, 2020 •

edited

willthefrog Aug 27, 2020

shippingwang Aug 27, 2020

willthefrog Aug 27, 2020

shippingwang Aug 27, 2020

willthefrog left a comment

shippingwang commented Aug 28, 2020 •

edited

willthefrog left a comment

XiaoguangHu01 left a comment

XiaoguangHu01 Aug 28, 2020

phlrain left a comment

		return int(np.ceil((index + 1) * input_size / output_size))


		def avg_pool1D_forward_naive(x,

[API2.0]Unify pooling function and add adaptive max pooling function #26483

[API2.0]Unify pooling function and add adaptive max pooling function #26483

Conversation

shippingwang commented Aug 20, 2020 • edited

PR types

PR changes

Describe

paddle-bot-old bot commented Aug 20, 2020

WuHaobo left a comment

Choose a reason for hiding this comment

willthefrog left a comment • 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

shippingwang Aug 27, 2020 • 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

willthefrog left a comment

Choose a reason for hiding this comment

shippingwang commented Aug 28, 2020 • edited

willthefrog left a comment

Choose a reason for hiding this comment

XiaoguangHu01 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

phlrain left a comment

Choose a reason for hiding this comment

shippingwang commented Aug 20, 2020 •

edited

willthefrog left a comment •

edited

shippingwang Aug 27, 2020 •

edited

shippingwang commented Aug 28, 2020 •

edited