Powf fix2 #1193
Powf fix2 #1193
Conversation
Codecov ReportAttention:
Additional details and impacted files@@ Coverage Diff @@
## main #1193 /- ##
==========================================
- Coverage 84.47% 84.44% -0.03%
==========================================
Files 548 549 1
Lines 61811 61905 94
==========================================
Hits 52216 52277 61
- Misses 9595 9628 33 ☔ View full report in Codecov by Sentry. |
| let pow = Data::from([[0.0, 1.0, 2.0], [3.0, 4.0, 5.0]]); | ||
| let tensor_pow = Tensor::<TestBackend, 2>::from_data(pow, &Default::default()); | ||
| let data_actual = tensor.powf(tensor_pow).into_data(); | ||
| let data_expected = Data::from([[1.0, 1.0, 4.0], [27.0, 256.0, 3125.0]]); |
There was a problem hiding this comment.
Regarding what you just said on discord,
So, I realized the issue with wgpu(and candle) might be effecting other test. assert_approx_eq seems to auto pass pairs if either one of them is NaN. looking into it its because any comparison with NaN is false, so the branch err>tol is never true
I think we should not test 0^0 = 1 here, because once we fix assert_approx_eq the test is gonna start failing. We could add a specific test for this after but it's complicated if it's not the same answer on all backend. And the right answer is debatable, I actually think NaN is truer than 1
Pull Request Template
Checklist
run-checks allscript has been executed.Related Issues/PRs
None
Changes
Fixes for powf wgpu function from pr #1184, splitting out so they can be merged
Testing
added a new test for powf under burn-tensor, first test has an element pair that confirms 0^0 works.
I realized candle wasn't being tested, and started testing it, encountered the same issue as before (0^0) being Nan instead of 1. I'm making an issue now, will link once I'm done