【Hackathon 6th No.30】为 paddle.nn.functional.max_pool1d/max_pool2d /max_pool3d/paddle.signal.stft 进行功能增强 #62975

AndPuQing · 2024-03-24T15:32:02Z

PR Category

User Experience

PR Types

New features

Description

修复了在return_mask=True模式下，ceil_mode失效的问题，并添加了相应单测。
~~修复了 stft 在 as_real 后出现的数据不一致问题~~(develop分支无法复现)

另外，stft 是否需要根据输入数据 x 所在的设备来决定输出结果的设备，develop分支中，即使 x 在 cpu 上，输出tensor 依然在gpu上(因为window 创建时是根据default_device决定的)

paddle-bot · 2024-03-24T15:32:07Z

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

zhwesky2010

max_pool1d/2d/3d重点需要检验下在ceil_mode为True/False下的shape是否一致？
paddle.signal.sftf与torch.sftf结果做了对比吗，详细对比两者下两者的结果
设备的问题：正常情况下，API的输入输出的设备是一致的，无需特殊指定配置，这里为何会出现不一致并且要手动指定？

zhwesky2010 · 2024-04-01T09:31:44Z

python/paddle/signal.py

@@ -362,6 +362,7 @@ def stft(
        ), f'expected a 1D window tensor of size equal to win_length({win_length}), but got window with shape {window.shape}.'
    else:
        window = paddle.ones(shape=(win_length,), dtype=x.dtype)
+        window = paddle.to_tensor(window, place=x.place)


这里windows的place为何是错的，一般情况下无需针对输出设备进行设置的

这里当用户默认设备是GPU，但是输入的X指定为CPU上，同时用户没有提供window输入，这里会创建一个在默认设备(GPU)上的tensor，最后造成输出时result在GPU，这是符合预期的吗？

zhwesky2010 · 2024-04-01T09:32:30Z

python/paddle/signal.py

@@ -362,6 +362,7 @@ def stft(
        ), f'expected a 1D window tensor of size equal to win_length({win_length}), but got window with shape {window.shape}.'
    else:
        window = paddle.ones(shape=(win_length,), dtype=x.dtype)
+        window = paddle.to_tensor(window, place=x.place)



正常情况下：API的输入与输出是在相同设备上的，无需指定，底层执行器会自动设定相应的place

test/legacy_test/test_pool1d_api.py

test/legacy_test/test_pool2d_api.py

test/legacy_test/test_pool3d_api.py

zhwesky2010 · 2024-04-01T09:40:51Z

@AndPuQing 另外，该PR合入后，需要提交 API中文文档、API映射文档的修改。

同时需要将PaConvert中 _test 屏蔽的case打开，提交一个PR会自动触发测试。

paddle-ci-bot · 2024-04-04T03:12:46Z

Sorry to inform you that 02eb1dc's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

AndPuQing · 2024-04-06T08:36:51Z

Paddle version

commit: 02eb1dc336cf650ac3b5c5d2fdbf55f893f7fb02
cuda: 12.0
cudnn: 8.9.1
nccl: 21701
xpu: False
xpu_xccl: False
xpu_xhpc: False
cinn: False

Pytorch version

2.2.2+cu121
39901f229520a5256505ec24782f716ee7ddc843

input dtype = float32

input_shape	n_fft	hop_length	normalized	onesided	Max absolute diff.	Max relative diff.
(2,100)	50	15	False	True	1.3126562e-06	9.709884e-06
(2,100)	100	15	False	True	3.8146973e-06	9.33037e-06
(2,500)	50	15	False	True	1.9073486e-06	9.538695e-06
(2,500)	100	15	False	True	2.1324806e-06	8.827294e-05
(2,500)	200	15	False	True	7.6293945e-06	8.2466126e-05
(2,500)	200	50	False	True	3.8146973e-06	5.94125e-05
(2,500)	200	15	True	True	4.7683716e-07	7.2988776e-05
(2,2000)	400	15	False	True	1.5258789e-05	0.00011829

input dtype = float64

input_shape	n_fft	hop_length	normalized	onesided	Max absolute diff.	Max relative diff.
(2,100)	50	15	False	True	3.55271368e-15	5.96877696e-15
(2,100)	100	15	False	True	7.10542736e-15	4.95725428e-14
(2,500)	50	15	False	True	3.55271368e-15	2.49609664e-14
(2,500)	100	15	False	True	7.10542736e-15	4.42488659e-14
(2,500)	200	15	False	True	1.42108547e-14	1.87328036e-13
(2,500)	200	50	False	True	9.28696749e-15	1.8887612e-13
(2,500)	200	15	True	True	1.33226763e-15	5.8703342e-14
(2,2000)	400	15	False	True	1.77635684e-15	2.74193926e-13

Test Code

import paddle.cuda_env
import paddle.version
import torch
import paddle
import numpy as np
import torch.version

paddle.version.show()
print(torch.version.__version__)
print(torch.version.git_version)
n_fft = 400
hop_length = 15
normalized = False
onesided = True
inputs_np = np.random.rand(2, 2000).astype('float32')
window_np = np.hamming(n_fft).astype('float32')

# Paddle
x = paddle.to_tensor(inputs_np)
window = paddle.to_tensor(window_np)
res = paddle.signal.stft(
    x=x,
    n_fft=n_fft,
    hop_length=hop_length,
    window=window,
    normalized=normalized,
    onesided=onesided,
)

res_a = paddle.as_real(res)

# PyTorch
x = torch.tensor(inputs_np)
window = torch.tensor(window_np)
res0 = torch.stft(
    input=x,
    n_fft=n_fft,
    hop_length=hop_length,
    window=window,
    normalized=normalized,
    onesided=onesided,
    return_complex=True,
)

res_b = torch.view_as_real(res0)

# Compare results
np.testing.assert_allclose(
    res.numpy(), res0.numpy(), rtol=1e-20, atol=1e-20, verbose=True
)

zhwesky2010 · 2024-04-08T07:50:49Z

@AndPuQing

第1个问题修复没问题
第2个问题验证无误，无需修改了
第3个问题，输出out与输入x是应该要保持place一致的，这个原因是paddle.to_tensor 默认会使用全局place，因此导致中间结果window的place与输入x不一致；这个问题需要修一下，另外这个文件中的其他window也有这个问题，统一修改下吧

luotao1 · 2024-04-08T09:16:32Z

第1个问题修复没问题

也可以单独提PR先把问题1给合入了 @AndPuQing

AndPuQing · 2024-04-12T14:30:40Z

@AndPuQing 另外，该PR合入后，需要提交 API中文文档、API映射文档的修改。

同时需要将PaConvert中 _test 屏蔽的case打开，提交一个PR会自动触发测试。

不知道这个文档指的什么文档呢？因为ceil_mode原本就有这个api.
问题3 我另外单独提PR

paddle-ci-bot · 2024-04-14T03:11:07Z

Sorry to inform you that c1bec1c's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

zhwesky2010 · 2024-04-17T06:40:12Z

@AndPuQing 另外，该PR合入后，需要提交 API中文文档、API映射文档的修改。
同时需要将PaConvert中 _test 屏蔽的case打开，提交一个PR会自动触发测试。

不知道这个文档指的什么文档呢？因为ceil_mode原本就有这个api.

问题3 我另外单独提PR

@AndPuQing

API中文文档需要修改：https://github.com/PaddlePaddle/docs 你修改了API，对应的API介绍文档，例如参数介绍这些都需要对应修改到与Paddle repo一致
API映射文档需要修改：https://github.com/PaddlePaddle/docs/tree/develop/docs/guides/model_convert/convert_from_pytorch/api_difference 找到对应的Pytorch API，API参数发生变化就需要修改对应的映射关系，比如与torch参数之间的对应关系发生变化

… fix-maxpool

…ndex

zhwesky2010

在 paddle/phi/ops/yaml/op_version.yaml 里需要修改下

action:
add_attr

zhwesky2010

LGTM

AndPuQing added 2 commits March 24, 2024 15:25

fix max_pool

a2b2974

fix stft

3b188b5

paddle-bot bot added the contributor External developers label Mar 24, 2024

Add ceil_mode parameter to MaxPool2d and MaxPool3d functions

a5a1227

luotao1 mentioned this pull request Mar 25, 2024

【Hackathon 6th】开源贡献个人挑战赛 #62905

Closed

Add test case for converting STFT result to real numbers

0c92d8e

luotao1 added the PaddlePaddle Hackathon label Mar 25, 2024

luotao1 assigned luotao1 and zhwesky2010 Mar 25, 2024

Fix adaptive pooling ceil_mode bug

02eb1dc

zhwesky2010 reviewed Apr 1, 2024

View reviewed changes

AndPuQing added 2 commits April 6, 2024 08:38

remove window to x device

f5c8ad4

add shape check for pool op

c1bec1c

luotao1 added the API label Apr 24, 2024

AndPuQing added 2 commits June 17, 2024 12:19

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

bd69130

… fix-maxpool

Add ceil_mode argument to max_pool2d_with_index and max_pool3d_with_i…

0463113

…ndex

zhwesky2010 previously approved these changes Jun 27, 2024

View reviewed changes

zhwesky2010 reviewed Jun 27, 2024

View reviewed changes

Add ceil_mode arg to max_pool2d_with_index and max_pool3d_with_index

1ebbd24

AndPuQing dismissed zhwesky2010’s stale review via 1ebbd24 July 1, 2024 04:02

zhwesky2010 approved these changes Jul 1, 2024

View reviewed changes

luotao1 assigned zyfncg and yuanlehome Jul 1, 2024

yuanlehome approved these changes Jul 1, 2024

View reviewed changes

zyfncg approved these changes Jul 2, 2024

View reviewed changes

luotao1 merged commit bdb15e3 into PaddlePaddle:develop Jul 2, 2024
31 of 32 checks passed

AndPuQing deleted the fix-maxpool branch July 12, 2024 09:06

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

【Hackathon 6th No.30】为 paddle.nn.functional.max_pool1d/max_pool2d /max_pool3d/paddle.signal.stft 进行功能增强 #62975

【Hackathon 6th No.30】为 paddle.nn.functional.max_pool1d/max_pool2d /max_pool3d/paddle.signal.stft 进行功能增强 #62975

AndPuQing commented Mar 24, 2024 •

edited

Loading

paddle-bot bot commented Mar 24, 2024

zhwesky2010 left a comment •

edited

Loading

zhwesky2010 Apr 1, 2024

AndPuQing Apr 5, 2024

zhwesky2010 Apr 1, 2024

zhwesky2010 commented Apr 1, 2024

paddle-ci-bot bot commented Apr 4, 2024

AndPuQing commented Apr 6, 2024

zhwesky2010 commented Apr 8, 2024 •

edited

Loading

luotao1 commented Apr 8, 2024

AndPuQing commented Apr 12, 2024

paddle-ci-bot bot commented Apr 14, 2024

zhwesky2010 commented Apr 17, 2024 •

edited

Loading

zhwesky2010 left a comment

zhwesky2010 left a comment

【Hackathon 6th No.30】为 paddle.nn.functional.max_pool1d/max_pool2d /max_pool3d/paddle.signal.stft 进行功能增强 #62975

【Hackathon 6th No.30】为 paddle.nn.functional.max_pool1d/max_pool2d /max_pool3d/paddle.signal.stft 进行功能增强 #62975

Conversation

AndPuQing commented Mar 24, 2024 • edited Loading

PR Category

PR Types

Description

paddle-bot bot commented Mar 24, 2024

zhwesky2010 left a comment • edited Loading

Choose a reason for hiding this comment

zhwesky2010 Apr 1, 2024

Choose a reason for hiding this comment

AndPuQing Apr 5, 2024

Choose a reason for hiding this comment

zhwesky2010 Apr 1, 2024

Choose a reason for hiding this comment

zhwesky2010 commented Apr 1, 2024

paddle-ci-bot bot commented Apr 4, 2024

AndPuQing commented Apr 6, 2024

Paddle version

Pytorch version

Test Code

zhwesky2010 commented Apr 8, 2024 • edited Loading

luotao1 commented Apr 8, 2024

AndPuQing commented Apr 12, 2024

paddle-ci-bot bot commented Apr 14, 2024

zhwesky2010 commented Apr 17, 2024 • edited Loading

zhwesky2010 left a comment

Choose a reason for hiding this comment

zhwesky2010 left a comment

Choose a reason for hiding this comment

AndPuQing commented Mar 24, 2024 •

edited

Loading

zhwesky2010 left a comment •

edited

Loading

zhwesky2010 commented Apr 8, 2024 •

edited

Loading

zhwesky2010 commented Apr 17, 2024 •

edited

Loading