| Re: Automatic Binsize Calculations [message #76263 is a reply to message #76262] |
Sun, 29 May 2011 11:20   |
manodeep@gmail.com
Messages: 33
Registered: June 2006
|
Member |
|
|
On May 29, 11:42 am, David Fanning <n...@idlcoyote.com> wrote:
> Gianguido Cianci writes:
>> Here's what I came up with, using sshist_2d.pro
>> (http://tinyurl.com/3on7bzx) that automagically finds bin size:
>
> I don't have a television, so while I listened to Djokovic
> defeat Gasquet on the French Open Radio I was fooling
> around using the 1D version of sshist to calculate
> a default bin size for cgHistoplot. What I discovered
> is that I get completely different results depending
> on the data type of the input data!
>
> I modified sshist a bit to get the bin size out of it
> as a keyword:
>
> ; Author: Shigenobu Hirose at JAMSTEC
> ; based on original paper
> ; Shimazaki and Shinomoto, Neural Computation 19, 1503-1527, 2007
> ; http://toyoizumilab.brain.riken.jp/hideaki/res/histogram.htm l
> ;
> function sshist, data, x=x, cost=cost, nbin=nbin, binsize=binsize
>
> COMPILE_OPT idl2
>
> nbin_min = 2
> nbin_max = 200
>
> ntrial = nbin_max - nbin_min + 1
>
> nbin = INDGEN(ntrial) + nbin_min
>
> delta = FLTARR(ntrial)
> cost = FLTARR(ntrial)
>
> for n = 0, ntrial-1 do begin
> delta[n] = (MAX(data) - MIN(data)) / (nbin[n] - 1)
>
> k = HISTOGRAM(data, nbins=nbin[n])
>
> kmean = MEAN(k)
> kvari = MEAN((k - kmean)^2)
> cost[n] = (2. * kmean - kvari) / delta[n]^2
> endfor
>
> n = (WHERE(cost eq MIN(cost)))[0]
> k = HISTOGRAM(data, nbins=nbin[n], locations=x, reverse_indices=ri)
>
> if arg_present(binsize) then binsize = delta[n]
> return, k
>
> end
>
> But, look at this:
>
> IDL> void = sshist(cgdemodata(21), binsize=bs) & print, bs
> 9.00000
> IDL> void = sshist(fix(cgdemodata(21)), binsize=bs) & print, bs
> 1.00000
> IDL> void = sshist(long(cgdemodata(21)), binsize=bs) & print, bs
> 1.00000
> IDL> void = sshist(float(cgdemodata(21)), binsize=bs) & print, bs
> 1.33684
>
> I have NO idea why this is occurring. :-(
>
If I set the "x" keyword to sshist, I see that the range returned is:
(note, cgdemodata(21) returns a [432,389] byte array ranging between 1
and 255 for me)
byte : 0-255 [bs = 84 and not 9 like David has]
int : 1-147 [bs = 1]
long : 1-147 [bs = 1]
float: 1-255 [bs = 1.33]
There must be a data-type mismatch going on somewhere. Only the float
calculation returns the histogram for the actual data range.
If I change the delta[n] line in sshist to
delta[n] = (max(data) - min(data))/(nbin[n] - 1.0)
i.e., force the calculation to be in floating point, then the int/long
types also return the range 1-255 (with a binsize of 2.0). The byte
calculation still has the same range but bs changes to 84.66. Not
entirely sure I understand what is going on..
Cheers,
Manodeep
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
