A modern programming language that makes developers happier.

kotlinlang

<@U0BN58B38> is the goal with koma to add support for as many backends as possible or just enough to have a performant option for each platform?

I'm asking because I just started brainstorming some ideas for a multiplatform ann kotlin api that would support various backends for actual execution. So for example it could support dl4j on jvm, tensorflow on native, etc. but have a single kotlin api that wraps it all. I guess it's kind of similar to what keras does in Python. Also similar to what koma does for matrix operations I think (I just started trying out koma yesterday). A big part of this means wrapping ndimensional arrays and matrices which of course koma already does but only with a few backends. Specifically, I wanted to start expirimenting with dl4j which uses nd4j which koma doesn't support. I guess technically all that would be necessary would be a function to convert from a koma ndarray to a nd4j ndarray but this would have a perfance penalty. It might be what I do to start expirimenting.

I guess the immediate practical question is do you plan to support nd4j backend on jvm? And the more philisophical question is in general do you want to support a bunch of backends or just minimum number necessary for a multiplatform scientific computing library?