Saddle: Scala Data Library

http://saddle.github.com/

by aklein • 13 years ago

51 25 13 years ago

saddle.github.com

Saddle is a high-performance data manipulation library for Scala.

joshklein• 13 years ago

Congrats on the release. I can think of at least one big organization I've talked to that was chomping at the bit to try pandas but had too much of an existing commitment to Scala to take the Python plunge. [Disclaimer: brother of OP]

JPKab• 13 years ago

I love pandas, and I think this is going to be great.

Is there something like this for Clojure? I guess I'll have to pick up scala too. Coursera here I come.

draegtun• 13 years ago

>Is there something like this for Clojure?

Probably Incanter which uses the Parallel Colt Java library - http://incanter.org/ | https://sites.google.com/site/piotrwendykier/software/parall...

wiradikusuma• 13 years ago

How does it different than https://github.com/scalanlp/breeze?

aklein8• 13 years ago

Breeze is more targeted to NLP and machine learning. Saddle draws heavily on the design of pandas (python library) to provide data structures enabling "alignment-free programming". Saddle outsources nearly all its linear algebra and numerics capabilities.

shawnalaken• 13 years ago

lightning fast!

wheaties• 13 years ago

As one of the colleagues of the author of this library, I can give my semi-biased opinion. To be honest, having worked in a previous life with Numpy+SciPy, the appearance of Saddle in our tech stack made the reasoning of complex numerical code easier. I'd suggest using it not just for it's performance (quite impressive for a JVM based library) but more for it's clear API. Expressible code (clean code) is debugged faster and maintained with less overhead. This library will let your code become expressive as a numeric library can be without sacrificing some of the nicer language features you've come to rely upon (map, flatMap, etc.)

pathdependent• 13 years ago

Thank you!

Most of my colleagues do data analysis in Python given Numpy+SciPy. I like Python, but if possible, I'd rather do as much of my development in a single language, and I prefer Scala.

This library certainly does not replicate the extensive functionality offered in Python for data analysis, but it does have the potential to seed Scala development. I for one will be perusing the code this weekend, and picking an avenue for subsequent exploration.

aklein8• 13 years ago

Cool, I welcome the feedback!

achompas• 13 years ago

Congrats, Adam!

Do we have performance information yet, even on some basic, common use cases?

Also, the docs mention EJML as the backend for Saddle's data structures--do you have any thoughts on using EJML?

aklein8• 13 years ago

Thanks! I will do some follow-up posts on performance, but know that it has been a MAJOR design consideration.

Consider the following in Saddle:

  val s1 = Series(vec.rand(10000), Index(Vec(array.randIntPos(10000)) % 100))

  val s2 = Series(vec.rand(10000), Index(Vec(array.randIntPos(10000)) % 100))

  clock { s1.join(s2, how=index.OuterJoin) }

This clocks in at 19ms on my machine after Hotspot kicks in.

The equivalent pandas:

  In [10]: ix1 = np.random.random_integers(0, 100, 10000)

  In [11]: ix2 = np.random.random_integers(0, 100, 10000)

  In [12]: df1 = DataFrame({'x' : np.random.rand(10000)}, ix1)

  In [13]: df2 = DataFrame({'y' : np.random.rand(10000)}, ix2)

  In [14]: %timeit df1.join(df2, how='outer')

10 loops, best of 3: 37.7 ms per loop

aklein8• 13 years ago

PS Regarding EJML, after extensive research, I found it hands down the fastest pure-java implementation for doing linear algebra.

While it's maybe 2x-4x slower than JNI wrapped ATLAS or MKL, for the cases I deal with, it just doesn't matter vs ease of use.

That said, it's LGPL, so I made it easy to swap out for other matrix libraries if you need.

jfim• 13 years ago

It took me a while to realize there were implicit conversions in the companion objects that are necessary in order to get useful functionality out of the data structures.

It might be worth adding an example to make it a bit more explicit in the documentation, such as:

  import org.saddle.Vec._
  Vec(1,2,3).median // Returns 2

Other than that, it looks pretty cool, I'll go use it right now. :)

Edit: Formatting.

aklein8• 13 years ago

Edit: you want to

  import org.saddle._

to get all the implicit goodness. I'll add a note.

saintx3• 13 years ago

I was sort of sad to learn earlier this year that the scalala project had become inactive, and when a friend pointed me at Breeze, the first thing that concerned me was that it seemed to "do ALL the things!", rolling in a bunch of other functionality along with a scalala revamp. What I really wanted was an elegant, fast, well written numerical computing library in Scala, and this seems to be it. This is great. Now all we need is to be able to tell this to use GPU hardware acceleration under the hood for things like FFTs and we're set!

aklein8• 13 years ago

I'd love to explore GPU accelerated solutions. I need more hours in the day...

saintx3• 13 years ago

The more I look into this, the better it looks. In particular, I appreciate that you used good existing solutions on the backend (EJML, Apache commons math, etc) where appropriate.

wandermatt• 13 years ago

I didn't see sparse vector support. Assuming I didn't just overlook it, is it on the roadmap?

aklein8• 13 years ago

Depends what you mean by sparse vector support. Maybe what you're interested in is best served by Series:

  val s = Series(Vec(1,2,3), Index(0,5,10))

This gives you

  s: org.saddle.Series[Int, Int] = 
  [3 x 1]
  0  -> 1
  5  -> 2
  10 -> 3

Then, for instance,

  s(5,10)
  res0: org.saddle.Series[Int,Int] =
  [2 x 1]
  5  -> 2
  10 -> 3

MLnick• 13 years ago

For me at least, sparse vector support means you can do elementwise operations (on the non-sparse elements) and in particular linear algebra like vector dot-products and matrix-vector multiply.

saintx3• 13 years ago

By sparse vector support, do you mean the Sparse Fast Fourier Transform reported by MIT last year? http://www2.technologyreview.com/article/427676/a-faster-fou...

endofunctor2• 13 years ago

Great news! Are you planning to do any integration with Erik's spire (https://github.com/non/spire)? I believe, some libraries already started collaborating with it (https://github.com/twitter/algebird/issues/99 and https://github.com/typelevel/scalaz-contrib/tree/master/spir...).

aklein8• 13 years ago

I'm definitely interested in exploring Spire. I believe the functionality is almost entirely orthogonal.

endofunctor2• 13 years ago

Awesome, looking forward to using saddle in my next project!

bennylak• 13 years ago

great work! BIG like!

Saddle: Scala Data Library

Related Stories