Hi, I used it a little, and wrapped some part, maily for training and visualizing decision trees. It was nice. I do not yet an elegant, well thought API, though.

Great! Do you have the code in a public repo? Have you tested performance (running time, not accuracy)?

Hi, no public repo yet, and no performance tests. I'll try to put some repo with code examples in the next few days, but please don't expect anything exciting - these are mainly thin wrappers and some functions to traverse the trees and visualize them.

No worries, I just wanted to see a few examples, I'll try it myself, but I'm not great at Java interop, so I'd like to see some example before trying 😄

👍 I

I wrapped several interpolations and some statistics in fastmath library. Currently working on clustering. I found that some stuff was slower than Apache Commons Math versions (correlations as far as I remember)

https://generateme.github.io/fastmath/fastmath.interpolation.html

https://generateme.github.io/fastmath/fastmath.stats.html

regarding running time, it is 2-10times faster than kixi.stats, even counting transferring data from seqs to arrays. I've tested combined descriptive statistics on 1e7 samples. fastmath.stats/stats-map vs similar combination in kixi (with transduce/fuse)

@UAE4G28HX

@U1EP3BZ3Q this is interesting, I'm somewhat testing Smile vs scikit-learn and I found that is slower

Have you thought about using https://github.com/uncomplicate/neanderthal?

not yet, I don't need linear algebra now

I believe that scikit-learn can be faster, it's partly implemented with cython. Smile is pure Java. Which is usual faster than pure Clojure.

http://haifengl.github.io/smile/images/benchm-ml.png so they are basically lying?

probably not here, they tested some cases on some data, possibly the other cases with other data could give different results

comparing speed should be very rigorous, I don't know how it was made with SMILE.

check the disccussions about Neanderthal vs ND4J speed here https://dragan.rocks/articles/18/Neanderthal-vs-ND4J-vol1 it's also the story about speed measurement traps

I'm actually working on that, I'm doing numpy vs Neanderthal

I know there are issues, that's why I'm running my benchmarks with workflows similar to what I need