Dear Falk,
On 31 Jan 2018, at 13:50, Falk Lüsebrink wrote:
> Dear Christian and Guillaume,
>
> thanks for the quick reply and thanks for digging into the issue.
>
> Is it be possible to include an option for single precision in the
> expert options? This could allow me to split the job into separate
> processes without running out of memory.
No, we already use single precision floats internally if it’s
possible. You can try the new CAT12 version r1265 which is slightly less
memory demanding because I have replaced some functions. However, during
LAS and some other steps there are still memory peaks of 50-60GB which
is not too surprising for such a challenging dataset with 500 million
voxels. I have successfully preprocessed the dataset on my 32GB machine,
so you may try to run 2 or even 3 processes on your machine at once.
Anyway, the SNR of these 0.25mm data is probably too low to use that
high spatial resolution and you should think about to resample the data
to 0.35-0.5mm with slight loss in spatial resolution, but the better SNR
might lead to better overall quality.
Best,
Christian
>
> Best,
> Falk
>
> -----Ursprüngliche Nachricht-----
> Von: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
> Im Auftrag von Christian Gaser
> Gesendet: Mittwoch, 31. Januar 2018 09:59
> An: [log in to unmask]
> Betreff: Re: [SPM] CAT12: Segmentation fault while processing high
> resolution dataset
>
> Dear Guillaume,
>
> it's exactly as you suspect and the crash occurs during calling
> spm_conv_vol. The data are challenging:
> vol 642x882x882 3995417664 double
>
> Changing the data to single was successful, so I try to optimize that
> part of the code to be not that memory demanding.
>
> Best,
>
> Christian
>
> On Tue, 30 Jan 2018 17:39:54 +0000, Guillaume Flandin
> <[log in to unmask]> wrote:
>
>> Dear Falk,
>>
>> Let's wait for Christian's feedback if this issue is CAT12 specific
>> but
>> I'm curious because the segmentation fault takes place in
>> spm_conv_vol.
>> From what you are saying, it seems that there is a minimal voxel size
>> under which the crash occurs so it could relate to file size or
>> numeric
>> overflow. Would you be able to stop CAT12 just before the call to
>> spm_conv_vol during "SPM preprocessing 2 (write)" and report what the
>> sizes of the input variables are?
>>
>> Best regards,
>> Guillaume.
>>
>>
>> On 29/01/18 13:38, Falk L sebrink wrote:
>>> Dear all,
>>>
>>>
>>>
>>> I m trying to process a high resolution dataset with 0.25 mm
>>> isotropic resolution using CAT12 (r1258) and SPM12 (v7219). It used
>>> to work with older versions, however, now it crashes during the SPM
>>> preprocessing 2 (write). See attached log file. Downsampling the
>>> dataset to a resolution worse or equal to 0.4 mm works fine, i.e.
>>> 0.33 mm does not work either.
>>> Bias field correction and such of SPM12 work without issues even at
>>> full resolution. So I suppose it is unrelated to SPM12.
>>>
>>>
>>>
>>> RAM and disk space is available plentiful (128 GB and a few TB
>>> respectively). Changing parameters (i.e. bias field correction
>>> strength, sampling distance, etc.) did not seem to help either.
>>>
>>>
>>>
>>> Any help is appreciated!
>>>
>>>
>>>
>>> Best,
>>>
>>> Falk
>>>
>>
>> --
>> Guillaume Flandin, PhD
>> Wellcome Trust Centre for Neuroimaging
>> University College London
>> 12 Queen Square
>> London WC1N 3BG
|