drop numpy dependency from Python code for cases without vectors about m2cgen

[izeigerman]

Hey @StrikerRUS, sorry, didn't expect you to reply so quickly :) I've edited my previous message right when you posted your reply. I've read your post carefully one more time and came to a conclusion that what you're saying makes sense.

[izeigerman]

Hey @StrikerRUS! Thanks for creating this issue.

Besides using pow and exp we also perform vector additions and multiplications by scalar. This can be potentially replaced with manually written functions similar to ones we introduced in other languages. However I don't see much benefit of doing this, especially considering that numpy is basically a de facto standard in the Data Science and Machine Learning community.

UPD However I don't have strong arguments against using exp and pow from math package. If there is an evidence that they perform faster than numpy (and apparently there is) then we should consider using them instead.

[StrikerRUS]

@izeigerman Thanks a lot for your prompt response!

Besides using pow and exp we also perform vector additions and multiplications by scalar.

Is self.with_vectors indicator used for that? I'm not suggesting replacing vector operators with manually written functions. I just suggested dropping numpy for fully scalar operations.

However I don't see much benefit of doing this,

I guess ~8x speedup worth it!

especially considering that numpy is basically a de facto standard in the Data Science and Machine Learning community.

Sure, but generated runtime Python module can be used far away from data science community 🙂 .

[izeigerman]

Great research btw 👍

[StrikerRUS]

@izeigerman No problem! I'm sorry too: I read fast, type slowly and GitHub doesn't provide live updates for editions. 😃
I'm glad you liked the idea!