OP Test Input Shape Requirements (English Version)
zhupengyang edited this page Jan 19, 2020
·
5 revisions
Specification summary：
- Section 1, Explanations Of The Specification
- Section 2, Check Of The Specification
- Section 3, Check Of CI
- Section 4, Suggestions Of Check Failure
Supplementary note：
- You may find some aspects that are not taken into account in the existing specifications, which need to be continuously supplemented and improved during the implementation process. Please feel free to give your feedback.
There is a problem that the input shape is too small in several OP tests, which leads to incomplete test. The grad check of some kernels is related to the size of the input shape. If the input shape is too small, the grad check passes, but it may still have problems.
In order to better ensure the coverage of OP tests, this specification is proposed. The corresponding detections are added to Op tests, and CI will check the white list of this specification.
- Op tests shape rule: The product of the input dimension of all test cases must be greater than or equal to 100.
- The white list of this specification wil skip specific OPs with limited input shape and OPs to be repaired. White list path: python/paddle/fluid/tests/unittests/white_list/check_shape_white_list.py
- The whitelist will be checked in CI. If you need to modify it, please contact one of specific people for approval.
specific cases:
- Shape check will be skipped if the op test is decorated by skip_check_grad_ci.
- Please contact one of specific people for approval.
- enlarge input shapes of all cases to 100+
- Please use skip_check_grad_ci to skip the shape check for specific cases，and add "[skip shape check]" at the beginning of the reason. Refer to PR#22332
If you encounter other problems, please contact @zhupengyang。