Hi Jonathan
We have done the validation before releasing the code in FSL, so you should not worry too much unless you see a very obvious problem. For details, see Moises’ paper:
http://www.ncbi.nlm.nih.gov/pubmed/23658616
Comparing the two versions is not straightforward, as you expect differences: a) due to differences in GPU/CPU precision, b) most importantly due to the stochastic nature of bedpostx - running it twice on the same machine and data will yield different results.
The way we compared the two versions was:
a) Remove random components from both versions and test their deterministic behaviour which yielded differences down to machine precision.
b) Test that the two versions give the same distributions on average (see paper above).
Cheers
Stam
On 4 Mar 2014, at 10:00, Jonathan Berrebi <[log in to unmask]> wrote:
> Hi,
>
> I got the bedpostx_gpu to work on debian thanks to neuro.debian and I was expecting numerical differences between CPU and GPU computing and we did get some.
>
> We can see them by substracting the images and by looking at the histogram. It is rather difficult to understand the substraction (normalised or not) of the images. On the other hand the histograms look very similar.
>
> Is there a secure way to validate the GPU computed results?
>
> Thank you,
>
> Best,
>
> Jonathan
|