Accelerating Python Libraries with Numba

http://www.continuum.io/blog/numba_growcut

by jasonsoja • 13 years ago

51 30 13 years ago

I looked at Numba a few weeks ago and it looked very impressive. One question I had that I did not see addressed in Numba's documentation was whether Numba is generally applicable or if it is focused on mathematical computation? I understand that math loops might be most in need to speed-up, but it would still be nice to speed-up non-math-centric code (e.g. Django...).

kingkilr• 13 years ago

It is not general purpose.

onalark10• 13 years ago

yet. :)

tlarkworthy4• 13 years ago

I get 10ms to cluster and segment an image with tuned numpy and opencv. With better results. Indeed the loops regarding the inner patch convolution are the killer. I used fingmineig, corner detectors, to turn a pixel location into a vector of corners at different spatial scales. Those routines are performed in parralell on opencv so I got rid of the inner patch kernel and it all delegates to blas. No loops

nightski• 13 years ago

I'd say that yes, if you are just performing an operation that already exists in a common C library just use that. But I could see the advantages for an algorithm that did not already exist for which you wanted to express in Python.

OpenCV is great as a black box but its built in functions are not always sufficient or optimal.

onalark10• 13 years ago

I'm always interested in working with practical examples. Do you have some sample code I could look at?

tlarkworthy4• 13 years ago

EDIT: Formatting is a bit whacked. I have two consecutive frames of a sonar image (same size) as the input, you will have to exchange those parts but I ran it fine. There are more parameters in the "DEFAULT_PARAMS" dict as I copied this from a much larger program. I work in greyscale so that might actually be a problem for genralization.

EDIT2: deleted as the src has been truncated, http://pastebin.com/dPMsRF78

aschreyer3• 13 years ago

It is a bit unfortunate that you have to "unroll" your code to get the most out of Numba. Hopefully at some point it will be able to translate NumPy functions such as np.abs(X - y).sum(axis=1)) efficiently into LLVM. Those extreme performance improvements are misleading though; in my experience the speed up compared to concise, optimal NumPy code is more in the range 50-100%.

dbecker2• 13 years ago

I'm sure numba is an impressive product. But I keep seeing examples on the continuum blog with an obscene number of for loops, and correspondingly great speedups.

My only response is: "My code doesn't look like that."

I'd be more impressed if they didn't cherry pick code with a dozen for-loops, and instead showed a moderate speedup from more standard (and vectorized) code.

onalark10• 13 years ago

(I wrote this last blog post)

In general, we're looking for results that will get people excited, and in some communities, we can get fairly silly speedups over native Python, so we are still picking that low-hanging fruit.

However, it's also important to be tied in to real applications. I'd be happy to take a crack at applying numba to your code. Do you have a reasonably small example we could take a look at?

travisoliphant3• 13 years ago

Numba's original purpose was not to speed up concise, optimal NumPy code. The fact that it can actually do that is an extra bonus. Numba's purpose was and still remains best at allowing you to still write the "hard-to vectorize" algorithms that require for-loops without having to use C/C++ or Fortran.

If you never are resorting to using C/C++ or Fortran, then Numba is not necessarily going to be helpful for you.

I don't think we are trying to mislead. We are simply showing examples that illustrate what Numba can be used for. Of course your code does not have many for-loops in it because you never could write code like that and have it be fast in Python. The big-deal about Numba is that now you can. You don't have to rely on a C/C++ library that either write or get someone else to write for you. You can just use Python.

I agree with you about the "unrolling". I don't recommend adding for-loops when good vectorized expressions work. Keep using them as we will be able to do a better job of parallelizing that code in the future. However, when you can't vectorize and for-loops are all you have, Numba is great.

We have spent some time on optimizing NumPy expressions like you describe and some array expressions can be translated down to equivalent LLVM loops. But most of the effort in that direction has been happening in the Blaze project which is generalizing NumPy so it can deal with much larger data but still produce optimized code from those vector expressions. This actually requires a re-thinking of how NumPy is implemented from the ground up.

skierscott• 13 years ago

Are there any Python vs. Numba vs. Cython vs C examples?

onalark10• 13 years ago

It's not common to see C/Fortran comparisons because they can be a pain to write. I'm implementing a serial C version for next week's blog post, though.

travisoliphant3• 13 years ago

Yes, there are several. See slide 15 of this talk http://www.slideshare.net/teoliphant/numba-siam-2013 and also the github repo: https://github.com/teoliphant/speed

Note that the array-expressions previously in numbapro only have been moved to numba (our plan for all premium products is to move features into open source as we get funds to support that).

travisoliphant3• 13 years ago

But, the upshot is that Numba produces code that is either faster or roughly the same speed as C or Cython.

Accelerating Python Libraries with Numba

Related Stories