[API Compatibility] align the loss apis --part by Manfredss · Pull Request #78811 · PaddlePaddle/Paddle

Manfredss · 2026-04-27T01:18:10Z

PR Category

User Experience

PR Types

New features

Description

align the loss apis in nn.layer and nn.functional

是否引起精度变化

否

paddle-bot · 2026-04-27T01:18:16Z

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

这些没对齐的关键是 https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/utils/decorator_utils.py#L801

Pytorch支持size_average、reduce的用法，用PaConvert先测试看看。

zhwesky2010 · 2026-04-27T03:33:18Z

                np.testing.assert_allclose(fetches[0], fetches[i], rtol=1e-5)


+class TestL1LossAPI(unittest.TestCase):


测下这些用法：size_average、reduce

codecov-commenter · 2026-04-27T15:14:25Z

Codecov Report

❌ Patch coverage is 97.19626% with 3 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@1ec99a3). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
python/paddle/utils/decorator_utils.py	95.58%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop   #78811   +/-   ##
==========================================
  Coverage           ?   97.19%           
==========================================
  Files              ?        3           
  Lines              ?      107           
  Branches           ?        0           
==========================================
  Hits               ?      104           
  Misses             ?        3           
  Partials           ?        0

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

Manfredss · 2026-04-27T20:00:52Z

/re-run all-failed

zhwesky2010 · 2026-04-28T03:49:27Z

这些装饰器体系有些复杂了，尽可能把这些装饰器合起来，代码复用：

别名装饰器通过配置来解决，每个API只能有一个装饰器

func和layer的装饰器能否合并？

zhwesky2010 · 2026-04-28T03:52:02Z

+                    _consume_legacy_positional_args(cls_name, args, kwargs)
                )
-            ):
-                raise_deprecated_error(cls_name, reduce_val, size_avg_val)


看一看怎么在之前的基础上进行最小改动：只在之前的基础上设置下suggested，将raise_deprecated_error换个名字。

Manfredss · 2026-04-29T22:36:12Z

/re-run all-failed

… ApiEnhance_loss

Manfredss · 2026-05-02T22:58:13Z

Paddle 与 PyTorch 在 loss API 上的不对齐主要分 5 类：

(1) 已废弃的 size_average/reduce 参数 PyTorch 仍接受并翻译为 reduction=，Paddle 之前直接报错；
(2) PyTorch 的位置布局比 Paddle 多 size_average/reduce 两个位置，且后续参数顺序也不同（如 CrossEntropyLoss 在 reduction 之后还插了 soft_label/axis/use_softmax）；
(3) 参数名差异（target vs label、beta vs delta、eps vs epsilon、anchor vs input 等）；
(4) API名差异（multilabel_* vs multi_label_*）；
(5) smooth_l1_loss 默认公式不同（PyTorch 是非 Huber，Paddle is_huber=True 是 Huber）。

想法是把第 (1)(2)(3) 类一次性折叠到一个机制里：在 decorator_utils.py 中维护一张 PYTORCH_LOSS_LAYOUT 表，记录每个 loss API 位置布局；legacy_reduction_decorator 在入口处检查 size_average 槽是否为 bool/None，若是则判定为 PyTorch 风格调用，按 PyTorch 名字把所有位置参数转成 kwargs，再 pop 掉 size_average/reduce 翻译成 reduction= 并 raise DeprecationWarning；否则视作 Paddle 内部调用（如 layer.forward 调 functional），原样透传不动。这样位置歧义、内部调用冲突、参数重命名（交给下游的 param_one_alias）三个问题在同一处统一解决

第 (4) 类靠 paddle.nn.functional.__init__ 里 4 行模块级别名解决；第 (5) 类用一个 11 行的 smooth_l1_beta_compat 装饰器：检测到 beta 时同时把 is_huber 默认设成 False，保留 Paddle 原有 delta 调用方的 Huber 默认。配套地，把 PoissonNLLLoss、TripletMarginLoss 两个 layer 上 @param_one_alias(["epsilon", "eps"]) 与 @legacy_reduction_decorator 的顺序对调（前者改到内层），这样位置参数到关键字参数转换出来的 eps 才能被 param_one_alias 接力改名为 epsilon

zhwesky2010 · 2026-05-07T09:28:13Z

Paddle 与 PyTorch 在 loss API 上的不对齐主要分 5 类：

(1) 已废弃的 size_average/reduce 参数 PyTorch 仍接受并翻译为 reduction=，Paddle 之前直接报错；

(2) PyTorch 的位置布局比 Paddle 多 size_average/reduce 两个位置，且后续参数顺序也不同（如 CrossEntropyLoss 在 reduction 之后还插了 soft_label/axis/use_softmax）；

(3) 参数名差异（target vs label、beta vs delta、eps vs epsilon、anchor vs input 等）；

(4) API名差异（multilabel_* vs multi_label_*）；

(5) smooth_l1_loss 默认公式不同（PyTorch 是非 Huber，Paddle is_huber=True 是 Huber）。

想法是把第 (1)(2)(3) 类一次性折叠到一个机制里：在 decorator_utils.py 中维护一张 PYTORCH_LOSS_LAYOUT 表，记录每个 loss API 位置布局；legacy_reduction_decorator 在入口处检查 size_average 槽是否为 bool/None，若是则判定为 PyTorch 风格调用，按 PyTorch 名字把所有位置参数转成 kwargs，再 pop 掉 size_average/reduce 翻译成 reduction= 并 raise DeprecationWarning；否则视作 Paddle 内部调用（如 layer.forward 调 functional），原样透传不动。这样位置歧义、内部调用冲突、参数重命名（交给下游的 param_one_alias）三个问题在同一处统一解决

第 (4) 类靠 paddle.nn.functional.__init__ 里 4 行模块级别名解决；第 (5) 类用一个 11 行的 smooth_l1_beta_compat 装饰器：检测到 beta 时同时把 is_huber 默认设成 False，保留 Paddle 原有 delta 调用方的 Huber 默认。配套地，把 PoissonNLLLoss、TripletMarginLoss 两个 layer 上 @param_one_alias(["epsilon", "eps"]) 与 @legacy_reduction_decorator 的顺序对调（前者改到内层），这样位置参数到关键字参数转换出来的 eps 才能被 param_one_alias 接力改名为 epsilon

方案基本没问题，可以再简化下：

第1~3类，给每个loss API维护一个别名表，也记录到LEGACY_POS中。对于关键字用法，则pop size_average/reduce，设置reduction，并修改别名。对于位置用法，由于目前的别名都是对应到了相同位置，应该别名不用做任何处理，而size_average/reduce直接去除，后面的参数位置参数也能对上，所以别名只需要删除size_average/reduce应该就可以？试试有无反例
第4类，处理好multi_label_的差异，然后新增API别名
第5类，单独处理，单独一个装饰器，参考index_add_decortor，看能否通过关键字与位置类型来区分出两者不同签名

…h<->Paddle 参数别名合并到一张 ``LEGACY_POS`` 表中，由 ``legacy_reduction_decorator`` 一次性完成位置 -> 关键字、关键字别名重命名、``size_average/reduce`` → ``reduction`` 的全部转换；同时移除，loss API 上多余的 ``@param_one_alias`` / ``@param_two_alias`` / ``@ParamAliasDecorator`` 装饰器

Manfredss · 2026-05-09T01:01:00Z

/re-run all-failed

… indexed slot only when the value is a bool

Manfredss · 2026-05-10T20:22:55Z

/re-run all-failed

Manfredss · 2026-05-11T06:15:46Z

/re-run all-failed

zhwesky2010

减少下注释吧。风格参考：
https://github.com/pytorch/pytorch/blob/main/CLAUDE.md#coding-style-guidelines

代码自身即为注释

zhwesky2010 · 2026-05-11T09:17:41Z

    return decorator


+# PyTorch positional layout for each loss API. Each entry maps the API


注释可以不用写太多（只写必要的注释），从代码可读性与简洁的角度来设计代码，参考

https://github.com/pytorch/pytorch/blob/main/CLAUDE.md#coding-style-guidelines

zhwesky2010 · 2026-05-11T11:47:45Z

+    return len(use_args) > idx and use_args[idx] in allowed
+
+
+def legacy_reduction_decorator(fn=None, *, is_method=True):


这个实现看起来也挺复杂的，那还是用第一版实现吧：

不用维护_SA0_RD1这些信息了，每个loss API维护一个args_list+kwargs_change，直接按args_list匹配然后pop掉size_average/reduce，设置reduction，修改别名即可。

相当于paconvert的逻辑搬过来

zhwesky2010 · 2026-05-11T11:49:03Z

    """

    @legacy_reduction_decorator
+    @param_one_alias(["epsilon", "eps"])


每个API只允许一个装饰器

align the loss apis

97fea6d

paddle-bot Bot added the contributor External developers label Apr 27, 2026

Manfredss mentioned this pull request Apr 27, 2026

[API Compatibility] enhance the tests for loss apis -part PaddlePaddle/PaConvert#874

Open

zhwesky2010 added the API Compatibility label Apr 27, 2026

zhwesky2010 reviewed Apr 27, 2026

View reviewed changes

Manfredss added 2 commits April 27, 2026 17:43

fully aligned the loss apis

cfaca13

resolve conflict

2f43fc7

zhwesky2010 reviewed Apr 28, 2026

View reviewed changes

rename the param and merge the three decorators into one

db661f7

Manfredss added 2 commits May 2, 2026 13:36

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

2c33e6a

… ApiEnhance_loss

reviewed and improved

c61ca65

Manfredss added 2 commits May 10, 2026 16:05

simplify logic, refer to kwargs and positional args

ecaf49d

kwargs are popped, positional values are fetched from the list at the…

a292f10

… indexed slot only when the value is a bool

zhwesky2010 reviewed May 11, 2026

View reviewed changes

Manfredss added 2 commits May 12, 2026 15:12

mimic the logic of paconvert, one decorator per api.

f6da7ae

fix

4b632fe

		np.testing.assert_allclose(fetches[0], fetches[i], rtol=1e-5)


		class TestL1LossAPI(unittest.TestCase):

		return decorator


		# PyTorch positional layout for each loss API. Each entry maps the API

		return len(use_args) > idx and use_args[idx] in allowed


		def legacy_reduction_decorator(fn=None, *, is_method=True):

Conversation

Manfredss commented Apr 27, 2026

PR Category

PR Types

Description

是否引起精度变化

Uh oh!

paddle-bot Bot commented Apr 27, 2026

Uh oh!

zhwesky2010 left a comment

Choose a reason for hiding this comment

Uh oh!

zhwesky2010 Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

codecov-commenter commented Apr 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Manfredss commented Apr 27, 2026

Uh oh!

zhwesky2010 Apr 28, 2026

Choose a reason for hiding this comment

Uh oh!

zhwesky2010 Apr 28, 2026 • 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!

Manfredss commented Apr 29, 2026

Uh oh!

Manfredss commented May 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

zhwesky2010 commented May 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Manfredss commented May 9, 2026

Uh oh!

Manfredss commented May 10, 2026

Uh oh!

Manfredss commented May 11, 2026

Uh oh!

zhwesky2010 left a comment • 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!

zhwesky2010 May 11, 2026

Choose a reason for hiding this comment

Uh oh!

zhwesky2010 May 11, 2026

Choose a reason for hiding this comment

Uh oh!

zhwesky2010 May 11, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

codecov-commenter commented Apr 27, 2026 •

edited

Loading

zhwesky2010 Apr 28, 2026 •

edited

Loading

Manfredss commented May 2, 2026 •

edited

Loading

zhwesky2010 commented May 7, 2026 •

edited

Loading

zhwesky2010 left a comment •

edited

Loading