Predict and Scoring Bug in sklearnex.svm.SVR and Train_test_split index bug

[nilslacroix]

Describe the bug
When I use train_test_split the indeces are shuffled sometimes, sometimes not. Also if you use the scoring method of an SVR estimator multiple times in a row with the same estimator and the same parameters, it can produce different scores.

To Reproduce
well you should be able to write a simple reproducer yourself, since this occurs frequently.

Expected behavior
No bugs.

Environment:

• Windows 10
• Anaconda

[nilslacroix]

sklearnex.svm.SVR is also buggy. Using .predict() on the same X_test set in a pipeline with preprocessors and encoders yiels different predictions. Just do it like ~50 times in a row and you will see what I mean.

This is also reproducible with basic examples like:

https://intel.github.io/scikit-learn-intelex/samples/nusvr_medical_charges.html

Maybe this happens because I use AMD CPU or something.

[nilslacroix]

I confirmed this bug happens (at least the scoring) when using AMD Hardware. Please at least implement a warning message, which checks the current CPU so people get a notice that this library will only work with intel specific hardware.

[jamescporter]

any update on this?

[isuruf]

I looked into this and the issue seems to be because on AMD, the code is throttled to the sse2 baseline which seems to be buggy. uxlfoundation/oneDAL#3026 should indirectly solve this as AMD hardware newer than Bulldozer will use the sse42 codepath (or avx2 codepath) with that.