Hacker News Reader

Comment by nl | Hacker News Reader

I like Go a lot. I just wish there was as much support in Go for machine learning as there is in Python.

At the moment, I feel like the whole "Python is the second best language for every problem" makes it the default choice for a lot of people.

View on HN

I wrote a lot of (and still occasionally maintain) a library called golearn [1] which tried to bring first-class support for the “traditional” ML algorithms, but it seems that Go’s compiler and toolchain just don’t optimise well enough for Go to be very competitive on performance for that application. That’s not a criticism of Go, it’s just that C/C++/FORTRAN have decades of optimising compiler support and most competing solutions are Python sugar over a nest of such C/C++/FORTRAN libraries. Go’s also concurrent, which is nice, but we found cases where being too concurrent also cost performance for most people. I’m still optimistic that Go has a place in ML, but it’s probably not right now.

[1] https://github.com/sjwitworth/golearn

I do wonder if bridging to a good matrix/tensor library would be sufficient (as it is in Python).

I'm aware of the interesting work happening in Swift/Tensorflow and was wondering what the equivalent in Go would look like.

Have a look at gorgonia.

https://github.com/gorgonia/gorgonia

To add to that, while Gorgonia's Tensor (https://github.com/gorgonia/tensor) library is fairly mature, plenty of work can still be done. Gorgonia proper too has plenty of work to be done on improving the various VMs. Last, there is greenfield work to be done in making Gorgonia user friendlier

Proper link:

https://github.com/sjwhitworth/golearn

Apologies (I was on mobile!)

Reply on news.ycombinator.com