Hi, Bob, The problem is that I want to resample my data with another sampling rate. the two rates is very close. I use the formula: s(t)=sum(a_k*sinc(t-kTs)). the new sampling rate is Ts', so I have s(nTs')=sum(a_k*sinc(nTs'-kTs)). The sum index k is over the (-P, P), Centered at n. The n is start from zero. THe code is using two for loops and it is slow. The length of s(nTs) is very long, so it takes quite long time to do it. Thanks Frank> Date: Sun, 26 Oct 2008 22:45:42 -0500> From: robert.kern@gmail.com> To: numpy-discussion@scipy.org> Subject: Re: [Numpy-discussion] help to speed up the python code> > On Fri, Oct 24, 2008 at 11:30, frank wang wrote:> > Hi,> >> > I have to send this request second time since my first message contains the> > attached data file which is too big and was blocked by the system. So this> > time I will not attach the data file.> >> > I have converted a matlab function to python using numpy. both matlab and> > python run slow. I know that numpy has a lot of features, so I hope some> > experts can help me to speed up the code.> > Can you describe in higher level terms what you are trying to do? I'm> having trouble following the code.> > -- > Robert Kern> > "I have come to believe that the whole world is an enigma, a harmless> enigma that is made terrible by our own mad attempt to interpret it as> though it had an underlying truth."> -- Umberto Eco> _______________________________________________> Numpy-discussion mailing list> Numpy-discussion@scipy.org> http://projects.scipy.org/mailman/listinfo/numpy-discussion _________________________________________________________________ Stay organized with simple drag and drop from Windows Live Hotmail. http://windowslive.com/Explore/hotmail?ocid=TXT_TAGLM_WL_hotmail_102008

frank wang wrote:

Hi, Bob,

The problem is that I want to resample my data with another sampling rate. the two rates is very close. I use the formula:

s(t)=sum(a_k*sinc(t-kTs)).

the new sampling rate is Ts', so I have s(nTs')=sum(a_k*sinc(nTs'-kTs)). The sum index k is over the (-P, P), Centered at n. The n is start from zero. THe code is using two for loops and it is slow. The length of s(nTs) is very long, so it takes quite long time to do it.

Additionally, assuming you still want to do it yourself in python, the above is a convolution, and both numpy and matlab have 'fast' (e.g. implemented in C) implementation of this; see numpy.convolve. You can have much faster speed than O(M * P) where M is the length of the original signal using FFT, but if either the sinc 'filter' or your signal is large, it is more difficult to do it without using large amount of memory (you have to split the input signal, and even the filter impulse response if it is large, and use overlapp-kind of techniques, see http://ccrma.stanford.edu/~jos/sasp/Example_Overlap_Add_Convolution.html for an example in matlab). cheers, David

My understanding of the root of the problem is that you end up with doing many evaluations of sinc. If this is so, one suggestion is to go with pre-computed filters. For example, if you are resampling from 9 points to 10, essentially you're trying to go from a function that is defined on points 0, 1, 2..., 8 to a function defined on points 0, 0.9, 1.8, 2.7, ... 0.8. If you take point 2.7, it falls in between 2 and 3, and the value for your new function can be approximately computed as the weighted average: f(2.7) = 0.7*f(3) + 0.3*f(2). This is equivalent to convolving with a filter [0.3, 0.7]. The only problem is that the filter will vary from point to point. So, perhaps, you can pre-compute several filters for each type of new point (there will be 10 point types for this specific case). Your filters would also be larger in length, so that your interpolation is more precise. Hope this helps. On 10/31/08, Charles R Harris wrote:

On Thu, Oct 30, 2008 at 11:44 PM, frank wang wrote:

...

Hi, Bob,

The problem is that I want to resample my data with another sampling rate. the two rates is very close. I use the formula:

s(t)=sum(a_k*sinc(t-kTs)).

the new sampling rate is Ts', so I have s(nTs')=sum(a_k*sinc(nTs'-kTs)). The sum index k is over the (-P, P), Centered at n. The n is start from zero. THe code is using two for loops and it is slow. The length of s(nTs) is very long, so it takes quite long time to do it.

I think you can use some form of chirpz/Bluestein for this. Just think of the data as the Fourier transform of its spectrum.

Chuck

-- Not to laugh, not to lament, not to curse, but to understand. -- Spinoza