test_eigh_op, test_eigvalsh_op tolerance

[zlsh80826]

• 标题：test_eigh_op, test_eigvalsh_op 單測問題
• 版本、环境信息：
     1）PaddlePaddle版本：develop
     2）CPU/GPU：GPU, CUDA 11.6
     3）系统环境：Ubuntu20.04
     4）Python版本号: python3.8
• 复现信息:

ctest -R test_eigh_op
ctest -R test_eigvalsh_op

• 问题描述：请详细描述您的问题，同步贴出报错信息、日志/代码关键片段

1. 在 CUDA 11.6 後 cusolver 改進了 Ssyevd 的性能, 但同時和 numpy 的差異會略微放大 test_eigh_op, test_eigvalsh_op tolerance 需要修改至 2e-6 之上
2. Complex 單測需求 Hermitian matrix (對角元素 imag 的部分需要為 0), 參考文件, test_eigh_op, test_eigvalsh_op

[paddle-bot-old]

您好，我们已经收到了您的问题，会安排技术人员尽快解答您的问题，请耐心等待。请您再次检查是否提供了清晰的问题描述、复现代码、环境&版本、报错信息等。同时，您也可以通过查看官网API文档、常见问题、历史Issue、AI社区来寻求解答。祝您生活愉快～

Hi! We've received your issue and please be patient to get responded. We will arrange technicians to answer your questions as soon as possible. Please make sure that you have posted enough message to demo your request. You may also check out the API，FAQ，Github Issue and AI community to get the answer.Have a nice day!

[Zjq9409]

eigh 算子修改PR链接为：#39568

[zlsh80826]

Thank @Zjq9409. I am running the unit test on our environment to ensure the PR works well.
Hi @huangjun12, test_eigvalsh_op has same the issue with test_eigh_op. Could you help to update the test_eigval_sh_op.py? Thanks

[zlsh80826]

@Zjq9409 I have tested #39568, but it seems I had wrong conclusion in the past. The tolerance needs to be more higher. Following is the error message in our environment.
I investaged the reason and found that 2e-6 is for [5, 5] input(the case I investigated before). The [32, 32] input needs a higher tolerance (atol=6e-5, rtol=5e-6)

test_eigh_op failed
 .......F....
======================================================================
FAIL: test_check_output_gpu (test_eigh_op.TestEighGPUCase)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/opt/paddle/paddle/build/python/paddle/fluid/tests/unittests/test_eigh_op.py", line 75, in test_check_output_gpu
    np.testing.assert_allclose(
  File "/usr/local/lib/python3.8/dist-packages/numpy/testing/_private/utils.py", line 1530, in assert_allclose
    assert_array_compare(compare, actual, desired, err_msg=str(err_msg),
  File "/usr/local/lib/python3.8/dist-packages/numpy/testing/_private/utils.py", line 844, in assert_array_compare
    raise AssertionError(msg)
AssertionError: 
Not equal to tolerance rtol=2e-06, atol=2e-06

Mismatched elements: 10 / 32 (31.2%)
Max absolute difference: 5.054474e-05
Max relative difference: 4.041603e-06
 x: array([-2.78129 , -2.573316, -2.460475, -2.090341, -1.802978, -1.725366,
       -1.633694, -1.468698, -1.250094, -0.896905, -0.634157, -0.491122,
       -0.47392 , -0.3968  , -0.152721, -0.108726, -0.039511,  0.275858,...
 y: array([-2.781283, -2.573307, -2.460468, -2.090336, -1.802972, -1.72536 ,
       -1.633688, -1.468692, -1.250091, -0.896903, -0.634154, -0.49112 ,
       -0.473918, -0.396798, -0.15272 , -0.108726, -0.039511,  0.275857,...

----------------------------------------------------------------------
Ran 12 tests in 2.927s

FAILED (failures=1)

[Zjq9409]

@Zjq9409 I have tested #39568, but it seems I had wrong conclusion in the past. The tolerance needs to be more higher. Following is the error message in our environment. I investaged the reason and found that 2e-6 is for [5, 5] input(the case I investigated before). The [32, 32] input needs a higher tolerance (atol=6e-5, rtol=5e-6)

test_eigh_op failed
 .......F....
======================================================================
FAIL: test_check_output_gpu (test_eigh_op.TestEighGPUCase)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/opt/paddle/paddle/build/python/paddle/fluid/tests/unittests/test_eigh_op.py", line 75, in test_check_output_gpu
    np.testing.assert_allclose(
  File "/usr/local/lib/python3.8/dist-packages/numpy/testing/_private/utils.py", line 1530, in assert_allclose
    assert_array_compare(compare, actual, desired, err_msg=str(err_msg),
  File "/usr/local/lib/python3.8/dist-packages/numpy/testing/_private/utils.py", line 844, in assert_array_compare
    raise AssertionError(msg)
AssertionError: 
Not equal to tolerance rtol=2e-06, atol=2e-06

Mismatched elements: 10 / 32 (31.2%)
Max absolute difference: 5.054474e-05
Max relative difference: 4.041603e-06
 x: array([-2.78129 , -2.573316, -2.460475, -2.090341, -1.802978, -1.725366,
       -1.633694, -1.468698, -1.250094, -0.896905, -0.634157, -0.491122,
       -0.47392 , -0.3968  , -0.152721, -0.108726, -0.039511,  0.275858,...
 y: array([-2.781283, -2.573307, -2.460468, -2.090336, -1.802972, -1.72536 ,
       -1.633688, -1.468692, -1.250091, -0.896903, -0.634154, -0.49112 ,
       -0.473918, -0.396798, -0.15272 , -0.108726, -0.039511,  0.275857,...

----------------------------------------------------------------------
Ran 12 tests in 2.927s

FAILED (failures=1)

已经更新rtol和atol值

[huangjun12]

Thank @Zjq9409. I am running the unit test on our environment to ensure the PR works well.
Hi @huangjun12, test_eigvalsh_op has same the issue with test_eigh_op. Could you help to update the test_eigval_sh_op.py? Thanks

eigvalsh修复pr：#39841

[zlsh80826]

@huangjun12
The atol needs to be 6e-5 and rtol needs to be 5e-6. The reason is discussed at #39446 (comment).
Reference: #39568

[huangjun12]

@huangjun12
The atol needs to be 6e-5 and rtol needs to be 5e-6. The reason is discussed at #39446 (comment).
Reference: #39568

Done

[zlsh80826]

#40699

[zlsh80826]

This issue is blocked by #40699

[zlsh80826]

#40699 merged. Close this issue