MNT Remove fixtures for Numpy 1.3 and 1.4 compatibility #20323
Conversation
Pointers to `vec` and `mat` arrays have also been dropped. `vec` and `mat` are now memoryviews, and the state for pickling has been adapted to explicitely cast memoryview as numpy arrays to recover the same behavior.
|
Here are the result on those benchmark comparing pairwise distances computation on As reported, performances have decreased but for I've used a shared machine to perform the benchmarks. Though it was not overwhelmed with others tasks, I would be glad if someone could report their results using the benchmark script linked above as well. Full report → asv continuous -b DistanceBenchmark main remove-old-np-compat-fixtures
· Creating environments
· Discovering benchmarks.
·· Uninstalling from conda-py3.9-cython-joblib-numpy-scipy-threadpoolctl...
·· Installing d7ff1305 into conda-py3.9-cython-joblib-numpy-scipy-threadpoolctl..
· Running 14 total benchmarks (2 commits * 1 environments * 7 benchmarks)
[ 0.00%] · For scikit-learn commit d7ff1305 (round 1/1):
[ 0.00%] ·· Benchmarking conda-py3.9-cython-joblib-numpy-scipy-threadpoolctl
[ 7.14%] ··· dist_metrics.ChebyshevDistanceBenchmark.time_pairwise ok
[ 7.14%] ··· ======= ============= ============= ============
-- d
------- ----------------------------------------
n 5 10 100
======= ============= ============= ============
100 91.9±0.5μs 138±0.4μs 1.19±0ms
1000 8.02±0.03ms 13.2±0.05ms 115±0.2ms
10000 1.20±0.03s 1.75±0.03s 10.6±0.01s
======= ============= ============= ============
SOME BENCHMARKS HAVE CHANGED SIGNIFICANTLY. |
jjerphan
changed the title
Co-authored-by: Thomas J. Fan <[email protected]>
|
As of 125e4b6, performances have not changed (thanks @thomasjpfan, this was also the answer of the conversation). ASV Benchmark results |
glemaitre
left a comment
glemaitre
left a comment
There was a problem hiding this comment.
So I think that it is worth merging this then.
…#20323) Co-authored-by: Thomas J. Fan <[email protected]>
…#20323) Co-authored-by: Thomas J. Fan <[email protected]>
4 participants
Reference Issues/PRs
None.
What does this implement/fix? Explain your changes.
Some fixtures for Numpy 1.3 and Numpy 1.4 were included.
They are no longer needed.
Pointers to
vecandmatarrays have also been dropped.vecandmatare now memoryviews, and the statefor pickling has been adapted to explicitely cast
memoryview as numpy arrays to recover the same
behavior.
Any other comments?