[SciPy-user] Re: vectorize(function)/arraymap did not return arrays?

[Yichun Wei]

Hi Expects,

Sorry if I changed what I've said. The list comprehension could not work 
as fast as I expected... it does look like syntaxic suger when I use it 
in my application. This is my problem: basically what I need to do is,

res = [convolve1d(a[x,y], k[x,y]) for x in range(16) for y in range(16)]
res = sum(res)

k.shape is (16,16,40), a.shape is (16,16,1800) or so. So the result will 
be a 1-d array of length 1839.

I had to do this quickly for it will be repeated for more than 10000 
times in optimization. I implemented this in matlab and find it runs 
fairly quick  (conv in matlab is really quick. And I think matlab has 
been optimized for for-loops.)

After playing with convolution of scipy and for loops for several days, 
I realized that I have to write some C++ code to accomplish this in 
python without a prominent performance headache. However, if I want to 
implement the whole nested loops in C++ using weave.inline, I have to 
throw out all the numeric vectorized syntax and rewrite the convolve1d 
function again in C++ (don't I? Would CXX helps some with this problem? 
I had to admit that I am reluctant to code in C++). Also all the 
materials I read suggest avoiding calling back python functions in weave.

Could you share some opinion on this? Thanks in advance.

- yichun




Yichun Wei wrote:
> 
> Well, I found list comprehension is actually fast enough.
> 
> res = [myconvolve(a[x,y], k[x,y]) for x in range(16) for y in range(16)]
> 
> which is much faster than
> 
> for x in range(16):
>     for y in range(16):
>         .....
> 
> This thread says that list comprehension could be comparable to map, or 
> even better if taking in to consideration the time calling helper 
> functions in map:
> http://mail.python.org/pipermail/python-list/2005-January/259690.html
> 
> I did not try list comprehension earlier because I roughly remember I 
> was told that list comprehension is syntaxic suger of for loops...
> 
>