【Hackathon No.4】为 Paddle 新增 masked_scatter API -part #59383

yangguohao · 2023-11-27T02:21:51Z

PR types

New features

PR changes

APIs

Description

add masked_scatter for Paddle

paddle-bot · 2023-11-27T02:21:56Z

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.

zoooo0820 · 2023-11-30T08:36:43Z

python/paddle/tensor/manipulation.py

+    Inplace version of ``masked_scatter`` API, the output Tensor will be inplaced with input ``x``.
+    Please refer to :ref:`api_paddle_masked_scatter`.
+    """
+    # make sure the dtype of x and source is the same


不必要的注释可移除掉

zoooo0820 · 2023-11-30T08:47:15Z

test/legacy_test/test_masked_scatter.py

+        self.dtype = "uint16"
+        self.value_shape = (300, 300)
+
+


实现代码中的几个有检查的错误场景，辛苦也补充下单测吧

zoooo0820 · 2023-11-30T08:52:00Z

python/paddle/tensor/manipulation.py

+
+    value = value.flatten()[mask_prefix].reshape(mask.shape)
+    return paddle.where(paddle.cast(mask, dtype="bool"), value, x)
+


实现逻辑上，masked_scatter 和 masked_scatter_两个API有不同的实现逻辑。是某些操作在静态图下不支持，又希望masked_scatter 使用同一套代码的原因吗。

理论上这组API的实现内部只应该有少量inplace / outplace API的差异，可以简单分析下动态图下当前两种方式的实现有性能差异吗。

是的，主要是 broadcast_to 的这个操作，在静态图下的 shape 如果是 None 或者 -1 之类的时候没办法正确的 broadcast，后面是模仿 paddle.where 中的 broadcast 操作。我感觉在动态图下的 inplace 操作直接使用 broadcast_to 可以避免多余的 Op 操作，性能上应该会快一些。后面我也可以测试一下看看有没有明显的差距

修改为两个 API 的逻辑一致，性能只有较为轻微的差异

zoooo0820

lgtm

jeff41404 · 2023-12-05T03:42:08Z

python/paddle/tensor/manipulation.py

+            >>> value = paddle.ones([2,4], dtype="float32")
+
+            >>> out = paddle.masked_scatter(x, mask, value)
+            >>> print(out.shape)


It's best to give the values of x, mask and out in example code to make it easier for users to understand

sunzhongkai588 · 2023-12-05T07:24:47Z

python/paddle/tensor/manipulation.py

+
+            >>> import paddle
+
+            >>> x = paddle.randn([2, 2])


估计得加 seed 固定输出，否则示例检查过不了，参考固定的输出优于随机

示例代码检查没过，log 如下，可能是因为 x = paddle.randn([2, 2]) 没有 print(x)，导致没有获得输出。代码示例部分也都注意一下这吧～可以本地跑一下试试

seed 固定的情况下 CI 和本地的结果不同，我这里又重新改了一次。

jeff41404

LGTM

luotao1 · 2023-12-06T03:40:01Z

可以提交对应的中文文档

luotao1 · 2023-12-08T03:07:39Z

2023-12-06 01:49:14 ++ python sampcd_processor.py --debug --mode cpu
2023-12-06 01:49:14 ----------------Check results--------------------
2023-12-06 01:49:14 >>> Sample code test capacity: {'cpu'}
2023-12-06 01:49:14 >>> 1 sample codes ran failed in env: {'cpu'}
2023-12-06 01:49:14 <DocTest(<modname?> paddle.Tensor.masked_scatter:1:0 ln 1)>, running time: 1.003s
2023-12-06 01:49:14 >>> Mistakes found in sample codes in env: {'cpu'}!
2023-12-06 01:49:14 >>> Please recheck the sample codes.

luotao1 · 2023-12-11T06:30:37Z

2023-12-11 13:13:17 The following tests FAILED:
2023-12-11 13:13:17 	1916 - test_masked_scatter (Failed)

yangguohao · 2023-12-11T08:08:45Z

2023-12-11 13:13:17 The following tests FAILED:
2023-12-11 13:13:17 	1916 - test_masked_scatter (Failed)

我再 rerun 下试试，我没有修改过 test_masked_scatter 的文件，之前是通过 CI 的。

luotao1 · 2023-12-11T08:26:07Z

test_masked_scatter 可能存在随机挂 https://xly.bce.baidu.com/paddlepaddle/paddle/newipipe/builds/6512?module=github/PaddlePaddle/Paddle&pipeline=PR-CI-Coverage&branch=pull/59383(develop) 看一下历史记录，有超过一半是单测失败的

sunzhongkai588

LGTM

luotao1 · 2023-12-13T04:11:40Z

请提交中文文档

zoooo0820 · 2024-01-03T07:17:27Z

test_masked_scatter 可能存在随机挂 https://xly.bce.baidu.com/paddlepaddle/paddle/newipipe/builds/6512?module=github/PaddlePaddle/Paddle&pipeline=PR-CI-Coverage&branch=pull/59383(develop) 看一下历史记录，有超过一半是单测失败的

@yangguohao 你好，因为这个PR添加的单测中暴露了些问题，能辛苦再花些时间跟进下吗？

问题1：test_numel_error 不能在静态图下通过
这个和之前讨论的一样，静态图下不能通过tensor的值去做assert（返回是个tensor而不是具体的true/false）,需要修改下现有的函数逻辑

问题2：部分单测存在随机挂，如TestMaskedScatterAPIBroadcast3
这个问题的原因尚不清晰，辛苦再排查一下，目前看起来只有静态图分支会出现问题。初步看起来，通过移除对mask的broadcast逻辑似乎可以解决问题，但可能影响后续处理value的逻辑。

Paddle/python/paddle/tensor/manipulation.py

Lines 4586 to 4588 in 99af9f7

    
           zeros_like_x = paddle.zeros_like(x, dtype=int) 
        
           mask = paddle.add(paddle.cast(mask, dtype="int"), zeros_like_x) 
        
           mask_prefix = paddle.clip(mask.cumsum() - 1, min=0)

问题2的复现步骤可以拷贝一份单测中test_static_graph的代码，配置复用TestMaskedScatterAPIBroadcast3，该问题看起来和np初始化的值无关，可以设定特定的x和mask值方便定位，循环执行1000次。

如果定位到问题2和其他API或者框架机制有关，可以反馈出来，我们再看看怎么解决该问题。

yangguohao · 2024-01-14T01:31:44Z

test_masked_scatter 可能存在随机挂 https://xly.bce.baidu.com/paddlepaddle/paddle/newipipe/builds/6512?module=github/PaddlePaddle/Paddle&pipeline=PR-CI-Coverage&branch=pull/59383(develop) 看一下历史记录，有超过一半是单测失败的

@yangguohao 你好，因为这个PR添加的单测中暴露了些问题，能辛苦再花些时间跟进下吗？

问题1：test_numel_error 不能在静态图下通过
这个和之前讨论的一样，静态图下不能通过tensor的值去做assert（返回是个tensor而不是具体的true/false）,需要修改下现有的函数逻辑

问题2：部分单测存在随机挂，如TestMaskedScatterAPIBroadcast3
这个问题的原因尚不清晰，辛苦再排查一下，目前看起来只有静态图分支会出现问题。初步看起来，通过移除对mask的broadcast逻辑似乎可以解决问题，但可能影响后续处理value的逻辑。

Paddle/python/paddle/tensor/manipulation.py

Lines 4586 to 4588 in 99af9f7

zeros_like_x = paddle.zeros_like(x, dtype=int)

mask = paddle.add(paddle.cast(mask, dtype="int"), zeros_like_x)

mask_prefix = paddle.clip(mask.cumsum() - 1, min=0)

问题2的复现步骤可以拷贝一份单测中test_static_graph的代码，配置复用TestMaskedScatterAPIBroadcast3，该问题看起来和np初始化的值无关，可以设定特定的x和mask值方便定位，循环执行1000次。

如果定位到问题2和其他API或者框架机制有关，可以反馈出来，我们再看看怎么解决该问题。

您好，我这里还有几个问题：

问题1，我尝试了一下没有什么很好的方法解决，在静态图下只能跳过这个检查。或者有什么建议可以完成这个功能

问题2我复现了错误，在固定了seed尝试了不同的shape后，还是没办法定位问题所在。最后尝试在 logical_not 的时候将 mask 先转为 bool dtype 问题不再出现。之后会提相关的 pr进行修改。

add masked_scatter

c484f02

paddle-bot bot added the contributor External developers label Nov 27, 2023

luotao1 mentioned this pull request Nov 27, 2023

【PaddlePaddle Hackathon 5th】开源贡献个人挑战赛 #57262

Open

yangguohao and others added 2 commits November 27, 2023 11:46

add test file

309e049

Merge branch 'develop' into masked_scatter

2c060b1

luotao1 assigned luotao1 and zoooo0820 Nov 29, 2023

zoooo0820 reviewed Nov 30, 2023

View reviewed changes

yangguohao and others added 2 commits December 1, 2023 14:28

fix

f1dcb9c

Merge branch 'develop' into masked_scatter

881af9b

zoooo0820 approved these changes Dec 4, 2023

View reviewed changes

yangguohao requested a review from zoooo0820 December 4, 2023 03:51

luotao1 assigned jeff41404 and sunzhongkai588 Dec 4, 2023

jeff41404 reviewed Dec 5, 2023

View reviewed changes

fix doc

ef21a94

sunzhongkai588 reviewed Dec 5, 2023

View reviewed changes

fix doc

43b65f3

yangguohao requested review from jeff41404 and sunzhongkai588 December 6, 2023 02:29

jeff41404 previously approved these changes Dec 6, 2023

View reviewed changes

luotao1 added the PaddlePaddle Hackathon label Dec 8, 2023

yangguohao dismissed jeff41404’s stale review via c6faf4a December 8, 2023 07:44

fix doc

6403dc4

yangguohao force-pushed the masked_scatter branch from c6faf4a to 6403dc4 Compare December 10, 2023 11:22

yangguohao and others added 2 commits December 12, 2023 12:36

Merge branch 'develop' into masked_scatter

53a5b32

fix

6f4485b

yangguohao force-pushed the masked_scatter branch from a47cc43 to 6f4485b Compare December 12, 2023 05:05

sunzhongkai588 approved these changes Dec 13, 2023

View reviewed changes

luotao1 changed the title ~~【Hackathon No.4】为 Paddle 新增 masked_scatter API~~ 【Hackathon No.4】为 Paddle 新增 masked_scatter API -part Dec 13, 2023

luotao1 merged commit 8d717f3 into PaddlePaddle:develop Dec 13, 2023

yangguohao mentioned this pull request Dec 19, 2023

【Hackathon 5th No.4】为 Paddle 新增 masked_scatter API PaddlePaddle/docs#6405

Merged

yangguohao mentioned this pull request Jan 15, 2024

fix masked_scatter #60835

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

【Hackathon No.4】为 Paddle 新增 masked_scatter API -part #59383

【Hackathon No.4】为 Paddle 新增 masked_scatter API -part #59383

yangguohao commented Nov 27, 2023

paddle-bot bot commented Nov 27, 2023

zoooo0820 Nov 30, 2023

yangguohao Dec 1, 2023

zoooo0820 Nov 30, 2023

yangguohao Dec 1, 2023

zoooo0820 Nov 30, 2023

yangguohao Nov 30, 2023

yangguohao Dec 1, 2023

zoooo0820 left a comment

jeff41404 Dec 5, 2023

yangguohao Dec 5, 2023

sunzhongkai588 Dec 5, 2023

yangguohao Dec 6, 2023

sunzhongkai588 Dec 6, 2023

yangguohao Dec 10, 2023

jeff41404 left a comment

luotao1 commented Dec 6, 2023

luotao1 commented Dec 8, 2023

luotao1 commented Dec 11, 2023

yangguohao commented Dec 11, 2023

luotao1 commented Dec 11, 2023

sunzhongkai588 left a comment

luotao1 commented Dec 13, 2023

zoooo0820 commented Jan 3, 2024

yangguohao commented Jan 14, 2024


		value = value.flatten()[mask_prefix].reshape(mask.shape)
		return paddle.where(paddle.cast(mask, dtype="bool"), value, x)

【Hackathon No.4】为 Paddle 新增 masked_scatter API -part #59383

【Hackathon No.4】为 Paddle 新增 masked_scatter API -part #59383

Conversation

yangguohao commented Nov 27, 2023

PR types

PR changes

Description

paddle-bot bot commented Nov 27, 2023

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

zoooo0820 left a 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

jeff41404 left a comment

Choose a reason for hiding this comment

luotao1 commented Dec 6, 2023

luotao1 commented Dec 8, 2023

luotao1 commented Dec 11, 2023

yangguohao commented Dec 11, 2023

luotao1 commented Dec 11, 2023

sunzhongkai588 left a comment

Choose a reason for hiding this comment

luotao1 commented Dec 13, 2023

zoooo0820 commented Jan 3, 2024

yangguohao commented Jan 14, 2024