On Mon, Jan 18, 2010 at 4:56 PM, Ivan Lazar Miljenovic wrote:

Did you know about hmatrix (available on Hackage) before you wrote this?

yes. hmatrix is equivalent to other parts of numpy. iirc hmatrix is a wrapper for algorithms from GSL+BLAS+LAPACK. numkell is only equivalent to numpy's core array type & functionality, without numpy's included linalg algorithms (which in Python are especially needed since coding them in Python will result in very slow code) hTensor seems to be more similar to numkell, as it also provides a multi-dimensional array type. however, if I understand correctly, hTensor is quite different. * numkell array's axes are part of their types. in hTensor those are only known in run-time. so numkell is more type-safe imho. * numkell's array zips are lazy/not-memoized by default. I may be wrong on this, but it seems that hTensor always creates in-memory arrays.

"yairchu@gmail.com" writes:

...

Hi Cafe,

I've created a numpy equivalent for Haskell. (Numpy is a python library for multi-dimensional arrays and operations on them)

Code at http://github.com/yairchu/numkell (not yet on hackage because it needs better names)

A numkell array is a pair of a function from integer inputs and a range for its inputs (size). This allows for easy memoizing into in-memory arrays, and additionally, numkell arrays also support useful operations like numpy's newaxis and folding axes away. As the "Array" name was already taken, numkell's array is currently called "Funk" (name suggestions very appreciated).

An example: Given an bunch of vectors as a 2d array, compute the distance between each pair of vectors

{-# LANGUAGE DeriveDataTypeable, GeneralizedNewtypeDeriving, TypeOperators #-}

import Data.HList import Data.NumKell import Data.Typeable

newtype PersonIdx = PersonIdx Int   deriving (Enum, Eq, Integral, Num, Ord, Real, Show, Typeable)

newtype FeatureIdx = FeatureIdx Int   deriving (Enum, Eq, Integral, Num, Ord, Real, Show, Typeable)

let personProps = fFromList [[5,3,2],[4,8,1],[2,6,9],[5,3,0]] :: Funk (HJust PersonIdx :*: HJust FeatureIdx :*: HNil) Double

...

personProps           FeatureIdx   0   1   2 PersonIdx          +   -   -   -         0          | 5.0 3.0 2.0         1          | 4.0 8.0 1.0         2          | 2.0 6.0 9.0         3          | 5.0 3.0 0.0

...

sumAxes (fmap (** 2) (liftF2 (-) (personProps !/ (SNewAxis .*. HNil)) (personProps !/ (SAll .*. SNewAxis .*. HNil)))) (TFalse .*. TFalse .*. TTrue .*. HNil)

          PersonIdx    0    1    2    3 PersonIdx         +    -    -    -    -         0         |  0.0 27.0 67.0  4.0         1         | 27.0  0.0 72.0 27.0         2         | 67.0 72.0  0.0 99.0         3         |  4.0 27.0 99.0  0.0

In Python the last line looks shorter:

...

...

...

((personProps[newaxis] - personProps[:,newAxis]) ** 2).sum(2)

Mostly due to Python's slicing syntax sugar. Still, numkell has one large benefit over numpy (apart from being for Haskell): With numpy this example creates a temporary 3d array in memory. In numkell the array is not allocated in memory unless "fMemo" is called.

If anyone has comments, suggestions, naming suggestions, complaints, etc, I would very much like to hear.

cheers, Yair _______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

-- Ivan Lazar Miljenovic Ivan.Miljenovic@gmail.com IvanMiljenovic.wordpress.com