Scikit Intelex not usable with AMD CPU

[nilslacroix]

I just found out that if you use scikit-learn-intelex with an amd cpu, the scoring of regressors like SVR is inconsistent over the same data and parameter configuration. You could argue it may be stupid to use an intel library with amd, but it states nowhere that you can not use amd cpus with it ...

I don't know if you should write a warning about using this library, but introducing hardware dependencies into scikit is a big nogo for me. Maybe someone will read this when he gets strange bugs like I did.

[jjerphan]

scikit-learn-intelex is another project managed by Intel.

You should report issues on their issue tracker instead.

[nilslacroix]

I already did. I just thought it would be good to spread the knowledge, since the package is designed for scikit specifically and never mentions that it is only working on Intel CPU properly. In the end there is a high chance people might end up in this forum rather than in intelex. But I completely understand your reasoning.

[benjamineruvieru]

I did end up on this forum, thank you

[Kasun-Pathirage]

@benjamineruvieru Me too. Spent more than 5 days trying to figure out why my gridsearchcv with svm was giving non-deterministic results despite setting np.random.seed(42), or, rng=np.random.RandomState(42) etc. Decided to just try turning off scikit-learn-intelex, and there it was. Pretty sad that it doesn't work as expected on AMD CPUs though. Using that library indeed speeds up computation quite a lot.

[ZSZYoung]

@benjamineruvieru Me too. Spent more than 5 days trying to figure out why my gridsearchcv with svm was giving non-deterministic results despite setting np.random.seed(42), or, rng=np.random.RandomState(42) etc. Decided to just try turning off scikit-learn-intelex, and there it was. Pretty sad that it doesn't work as expected on AMD CPUs though. Using that library indeed speeds up computation quite a lot.

sad

[glemaitre]

As mentioned by @jjerphan, we are not maintaining nor developing scikit-learn-intelx.
@nilslacroix I would be interested to know if you installed the package on purpose? As far that I know, conda is kind of raising a warning advertising scikit-learn-intelx but I never saw an automatic install thought.

[nilslacroix]

I did install it on purpose, but I am not aware of any warnings nor have I seen any. I just stumpled upon the library during my research and it promised performance gains of factor 50-100 out of the box, which is huge, so I implemented it. I would have had no problem if there was a warning during training like "hey you are using hardware which is not working with this library". Right know it just produces wrong scores and index/column splits.

[glemaitre]

I am not aware of any warnings nor have I seen any

I was only referring to the warning raised by conda advertising installing scikit-learn-intelx.
Then, I actually never used scikit-learn-intelx and I would expect train_test_split to be working without any issue, whatever hardware :)

[RnoldR]

When installing scikit-learn via conda on my AMD powered laptop I got the following message: Installed package of scikit-learn can be accelerated using scikit-learn-intelex, about the opposite of a warning. Only later I realised the name "intel" in the URL which is why I ended up here. Maybe omiting the "advice" for AMD CPU's?

Thanks for the warnings @nilslacroix and @jjerphan.

[ChowChairman]

Well, according to their page
https://www.intel.com/content/www/us/en/developer/tools/oneapi/scikit-learn.html

it says it supports "All CPUs with x86 architecture"

[adrinjalali]

As already mentioned, we are not maintaining that package on the main conda package, nor are we the ones giving advice to people to install that package. This is not a scikit-learn core issue, and you should take it to the people maintaining those packages.

We have always worried about people being confused with that package, but since we have no control over it, we can't remove that suggestion.

You should ALWAYS install the package from pypi or conda-forge if you want us to support what you have installed. The main channel is not us.