[Dy2St] Make RandomRotation static comparison deterministic

[ShigureNyako]

PR Category

Environment Adaptation

PR Types

Not User Facing

Description

This PR fixes a recurring Mac-CPU CI flake in test_transforms_static, observed while validating #78558.

The failing case is TestRandomRotation_expand_True, which compares dynamic and static outputs with np.testing.assert_almost_equal. The current test tries to make RandomRotation effectively deterministic by passing a very narrow degree range, but it still samples the angle from two different RNG paths:

• dynamic mode: random.uniform(...)
• static mode: paddle.uniform(...)

That means the test is not actually comparing the same rotation angle in the two modes. On macOS this sometimes shows up as a few mismatched edge pixels after expand=True, causing the test to fail on one run and pass on rerun.

To fix that, this PR keeps the change scoped to the test file:

• add a local deterministic FixedAngleRandomRotation helper in test/legacy_test/test_transforms_static.py
• use the exact same fixed angle for both TestRandomRotation and TestRandomRotation_expand_True

This does not hide a real kernel bug. It makes the test validate the intended contract more accurately: dynamic and static RandomRotation should match when given the same input and the same angle.

是否引起精度变化

否

[paddle-bot]

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.

[ShigureNyako]

@SigureMo

我先自 review 过这一版了，改动只落在 test/legacy_test/test_transforms_static.py。

结论：这次修的是 RandomRotation 静态/动态图对比单测本身的随机性来源不一致问题，不是去放宽断言阈值。

关键信息：

• python/paddle/vision/transforms/transforms.py:1824-1830 里，RandomRotation._get_param() 在动态图走 random.uniform(...)，在静态图走 paddle.uniform(...)
• 原测试虽然把角度范围收得很窄，但两边实际上还是会采样到不同角度；在 macOS 上 expand=True 时就会偶发成边缘像素不一致
• 这版在测试文件里加了一个本地 FixedAngleRandomRotation helper，只把采样角度固定住，仍然走 RandomRotation 的旋转逻辑和 expand=True 路径

所以它不会掩盖真实算子 bug；如果同一个输入、同一个角度下静态/动态图旋转结果真的不一致，这个测试还是会继续失败。

[ShigureNyako]

/re-run all-failed

[SigureMo]

LGTMeow