Implement tensordot and repeat
#95
Conversation
Signed-off-by: neNasko1 <[email protected]>
Co-authored-by: Aditya Goel <[email protected]>
xfails.txt
Outdated
| array_api_tests/test_manipulation_functions.py::test_repeat | ||
| # segmentation fault: due to onnxruntime not handling einsum with dim 0 | ||
| array_api_tests/test_linalg.py::test_tensordot |
There was a problem hiding this comment.
If these tests segfault then we should put them into skips.txt rather than here. That the array-api-tests are green nonetheless is an indication that something else is fishy. That feeling is confounded by the fact that they finish within 20s. Could you take a look, please?
tests/test_core.py
Outdated
| (np.arange(60).reshape(3, 4, 5), np.arange(60), None), | ||
| ], | ||
| ) | ||
| def test_repeat(a, repeats, axis, lazy_repeats): |
There was a problem hiding this comment.
What is "lazy" about the lazy_repeats case here? Even when repeats is int, it's immediately wrapped in ndx.asarray(repeats) in your implementation of repeat.
There was a problem hiding this comment.
For a test that exercises shape inference for repeat, ideally you should be using a lazy array and just building the ONNX model.
There was a problem hiding this comment.
The idea of the test is to test that if we actually supply the argument as a python integer, we wrap it correctly. I rewrote the test to be cleaner.
Signed-off-by: Atanas Dimitrov <[email protected]>
|
Thanks for merging into main! |
Signed-off-by: Atanas Dimitrov <[email protected]>
This PR adds implementations to tensordot and repeat.
Note they currently segfault on the latest onnx release, so the arrayapi tests are disabled. Coverage is provided instead on our side with some tests.