Remove as_tensor argument of set_tensors_from_ndarray_1d
#2615
Conversation
facebook-github-bot
added
the
CLA Signed
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #2615 +/- ##
==========================================
- Coverage 99.98% 99.98% -0.01%
==========================================
Files 196 196
Lines 17372 17358 -14
==========================================
- Hits 17370 17356 -14
Misses 2 2 ☔ View full report in Codecov by Sentry. |
|
@saitcakmak has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
botorch/optim/utils/numpy_utils.py
Outdated
| torch.as_tensor(vals, device=tnsr.device, dtype=tnsr.dtype) | ||
| .to(tnsr) | ||
| .view(tnsr.shape) | ||
| .to(tnsr) |
There was a problem hiding this comment.
It seems that these .to() are noops in this setup given that the tensor is initially constructed on the right device with the right dtype?
| torch.as_tensor(vals, device=tnsr.device, dtype=tnsr.dtype) | |
| .to(tnsr) | |
| .view(tnsr.shape) | |
| .to(tnsr) | |
| torch.as_tensor(vals, device=tnsr.device, dtype=tnsr.dtype) | |
| .view(tnsr.shape) |
There was a problem hiding this comment.
Actually, let's do the reshaping first and then use from_numpy: torch.from_numpy(vals.reshape(tnsr.shape)).to(tnsr). I did some lightweight benchmarking and this generally is a tad faster.
There was a problem hiding this comment.
The benchmarking explains the 22 minute gap between the two comments :D. I updated the diff directly rather than waiting for back & forth.
There was a problem hiding this comment.
Looks like this doesn't work if vals is a scalar. Reverting to the previous solution
There was a problem hiding this comment.
Since there have been change after this comment, is there anything I can still do to assist? In any case, thanks for already implementing the necessary changes :)
There was a problem hiding this comment.
I think the PR is ready as is. I'll land it after the internal checks pass. Thanks for the contribution!
There was a problem hiding this comment.
Glad that I could help :)
|
@saitcakmak has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@saitcakmak has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@saitcakmak merged this pull request in 81c16ff. |
4 participants
Motivation
This PR removes the
as_tensorargument of theset_tensors_from_ndarray_1dfunction innumpy_utils.py.The reason for this change is that the previous implementation
float32precision, this always transformed floats intofloat64precisionThis change was discussed previously with @Balandat and @esantorella in #2596 .
Have you read the Contributing Guidelines on pull requests?
Yes, I have read it.
Test Plan
I checked the format using
ufmt format .and verified that all tests are still running usingpytest -ra.Related PRs
This PR is related to #2597 and #2596, fixing a similar issue (
float32being broken by some internal calculations).