[Numpy-discussion] A numpy accumulator...
Sat Oct 3 02:26:49 CDT 2009
This idea was inspired by a discussion at SciPY, in which we spent a
LOT of time during the numpy tutorial talking about how to accumulate
values in an array when you don't know how big the array needs to be
when you start.
The "standard practice" is to accumulate in a python list, then convert
the final result into an array. This is a good idea because Python lists
are standard, well tested, efficient, etc.
However, as was pointed out in that lengthy discussion, if what you are
doing is accumulating is a whole bunch of numbers (ints, floats,
whatever), or particularly if you need to accumulate a data type that
plain python doesn't support, there is a lot of overhead involved: a
python float type is pretty heavyweight. If performance or memory use is
important, it might create issues. You can use and array.array, but it
doesn't support all numpy types, particularly custom dtypes.
I talked about this on the cython list (as someone asked how to do
accumulate in cython), and a few folks thought it would be useful, so I
put together a prototype.
What I have in mind is very simple. It would be:
- Only 1-d
- Support append() and extend() methods
- support indexing and slicing
- Support any valid numpy dtype
- which could even get you pseudo n-d arrays...
- maybe it would act like an array in other ways, I'm not so sure.
- ufuncs, etc.
It could take the place of using python lists/arrays when you really
want a numpy array, but don't know how big it will be until you've
The implementation I have now uses a regular numpy array as the
"buffer". The buffer is re-sized as needed with ndarray.resize(). I've
enclosed the class, a bunch of tests (This is the first time I've ever
really done test-driven development, though I wouldn't say that this is
a complete test suite).
A few notes about this implementation:
* the name of the class could be better, and so could some of the
* on further thought, I think it could handle n-d arrays, as long as
you only accumulated along the first index.
* It could use a bunch more methods
- deleting part of eh array
- probably anything supported by array.array would be good.
* Robert pointed me to the array.array implimentation to see how it
expands the buffer as you append. It did tricks to get it to grow fast
when the array is very small, then eventually to add about 1/16 of the
used array size to the buffer. I imagine that this would gets used
because you were likely to have a big array, so I didn't bother and
start with a buffer at 128 elements, then add 1/4 each time you need to
expand -- these are both tweakable attributes.
* I did a little simple profiling, and discovered that it's slower
than a python list by a factor of more than 2 (for accumulating python
ints, anyway). With a bit of experimentation, I think that's because of
a couple factors:
- an extra function call -- the append() method needs to then do an
assignemt to the buffer
- Object conversion -- python lists store python objects, so the
python int can jsut go right in there. with numpy, it needs to be
converted to a C int first -- a bit if extra overhead.
Christopher Barker, Ph.D.
Emergency Response Division
NOAA/NOS/OR&R (206) 526-6959 voice
7600 Sand Point Way NE (206) 526-6329 fax
Seattle, WA 98115 (206) 526-6317 main reception
More information about the NumPy-Discussion