[torchlib] Fix aten::diagonal #1755
Conversation
justinchuby
added
the
topic: torch_lib
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #1755 +/- ##
==========================================
- Coverage 75.07% 75.06% -0.01%
==========================================
Files 245 245
Lines 26443 26437 -6
Branches 4824 4822 -2
==========================================
- Hits 19852 19846 -6
Misses 5660 5660
Partials 931 931 ☔ View full report in Codecov by Sentry. |
Test Results 24 files ±0 24 suites ±0 3h 14m 26s ⏱️ -8s For more details on these failures, see this check. Results for commit 27cf0e0. ± Comparison against base commit a72f048. This pull request removes 2 tests.♻️ This comment has been updated with latest results. |
justinchuby
requested review from
titaiwangms,
xiaowuhu,
shubhambhokare1 and
fatcat-z
|
If possible, it'd better paste the error we saw into description so we know what kind of logic was fixed. |
Done, thanks. |
| neg_1 = op.Constant(value_ints=[-1]) | ||
| dim1_size = op.Reshape(op.Gather(op.Shape(self), dim1), neg_1) # row | ||
| dim2_size = op.Reshape(op.Gather(op.Shape(self), dim2), neg_1) # col | ||
| dim1_size = op.Shape(self, end=dim1, start=dim1 + 1) # row |
There was a problem hiding this comment.
Can dim1 be negative? Especially -1? Same for dim2.
There was a problem hiding this comment.
ah it can. Thanks for pointing this out!
There was a problem hiding this comment.
Reverted to use gather. Added a comment
There was a problem hiding this comment.
Hmm but its handled in
There was a problem hiding this comment.
In any case using Shape(start, end) does not work for some reason
Turn aten::diagonal as trace only and fix its logic by explicitly converting python constants to onnx constants. This was needed because the exporter logic was not handling the type conversion correctly (yet)