Closed yuval-harpaz closed 7 years ago
the solution only depends on noise_cov and data_cov which are parameters. You can therefore call lcmv_raw on smaller slices of data. It will be equivalent although slightly slower.
would that be good enough?
right, sorry, should have thought of that. in any case, it would be more 'open source' like if the weights are accessible. In SAM beamforming you always get the weights (but not the forward model and other stuff), and in fieldtrip you get them if you specify cfg.keepfilter='yes'. but this would be lower priority thanks
On 9 December 2015 at 22:36, Alexandre Gramfort notifications@github.com wrote:
the solution only depends on noise_cov and data_cov which are parameters. You can therefore call lcmv_raw on smaller slices of data. It will be equivalent although slightly slower.
would that be good enough?
— Reply to this email directly or view it on GitHub https://github.com/mne-tools/mne-python/issues/2685#issuecomment-163384636 .
Yuval
Dr .Harpaz
BIU MEG lab http://faculty.biu.ac.il/~goldsa/index.html
write us how you would like to us it (2-3 lines code snippet) and we'll see what we can do.
thx
I would like to have a nsourse x nchan np.ndarray so I could then multiply it by the data to get raw source timecourse (no depth normalization). I suggest to give it as a second output variable for lcmv_raw, but do which ever cheepest way for you ... thanks
On 10 December 2015 at 09:50, Alexandre Gramfort notifications@github.com wrote:
write us how you would like to us it (2-3 lines code snippet) and we'll see what we can do.
thx
— Reply to this email directly or view it on GitHub https://github.com/mne-tools/mne-python/issues/2685#issuecomment-163526531 .
Yuval
Dr .Harpaz
BIU MEG lab http://faculty.biu.ac.il/~goldsa/index.html
It would be a trivial edit to the code, but an API change. We could of course habe some return_weights parameter that equals to False by default. Option2: maintain a copy of the code privately and extract what you wish for debugging purposes. Option 3: private function that returns the weights. The you call that one for debugging your analysis.
On 10 Dec 2015, at 08:25, yuval-harpaz notifications@github.com wrote:
right, sorry, should have thought of that. in any case, it would be more 'open source' like if the weights are accessible. In SAM beamforming you always get the weights (but not the forward model and other stuff), and in fieldtrip you get them if you specify cfg.keepfilter='yes'. but this would be lower priority thanks
On 9 December 2015 at 22:36, Alexandre Gramfort notifications@github.com wrote:
the solution only depends on noise_cov and data_cov which are parameters. You can therefore call lcmv_raw on smaller slices of data. It will be equivalent although slightly slower.
would that be good enough?
— Reply to this email directly or view it on GitHub https://github.com/mne-tools/mne-python/issues/2685#issuecomment-163384636 .
Yuval
Dr .Harpaz
BIU MEG lab http://faculty.biu.ac.il/~goldsa/index.html — Reply to this email directly or view it on GitHub.
all works for me. As is, I am analysing 10min recording and it is going to take the whole day
On 10 December 2015 at 10:32, Denis A. Engemann notifications@github.com wrote:
It would be a trivial edit to the code, but an API change. We could of course habe some return_weights parameter that equals to False by default. Option2: maintain a copy of the code privately and extract what you wish for debugging purposes. Option 3: private function that returns the weights. The you call that one for debugging your analysis.
On 10 Dec 2015, at 08:25, yuval-harpaz notifications@github.com wrote:
right, sorry, should have thought of that. in any case, it would be more 'open source' like if the weights are accessible. In SAM beamforming you always get the weights (but not the forward model and other stuff), and in fieldtrip you get them if you specify cfg.keepfilter='yes'. but this would be lower priority thanks
On 9 December 2015 at 22:36, Alexandre Gramfort < notifications@github.com> wrote:
the solution only depends on noise_cov and data_cov which are parameters. You can therefore call lcmv_raw on smaller slices of data. It will be equivalent although slightly slower.
would that be good enough?
— Reply to this email directly or view it on GitHub < https://github.com/mne-tools/mne-python/issues/2685#issuecomment-163384636
.
Yuval
Dr .Harpaz
BIU MEG lab http://faculty.biu.ac.il/~goldsa/index.html
— Reply to this email directly or view it on GitHub.
— Reply to this email directly or view it on GitHub https://github.com/mne-tools/mne-python/issues/2685#issuecomment-163533064 .
Yuval
Dr .Harpaz
BIU MEG lab http://faculty.biu.ac.il/~goldsa/index.html
all works for me. As is, I am analysing 10min recording and it is going to take the whole day
lcmv_raw for 10min of data processed by chunks takes 1 day???
well, I got to segment 600 of 1819 (quarter seconds segments) and it is 11:30 here, it will take a few hours more I do it with this loop
Kur=np.zeros((fwd['nsource'])) for sampi in range(0,len(samps)): start=samps[sampi] stop=start+dur stc = mne.beamformer.lcmv_raw(raw, fwd, empty_cov, data_cov,start=start, stop=stop) kur=g2(stc.data) kur[kur<0]=0 Kur=Kur+kur print("DONE "+str(sampi+1)+" of "+str(len(samps)))
On 10 December 2015 at 11:07, Alexandre Gramfort notifications@github.com wrote:
all works for me. As is, I am analysing 10min recording and it is going to take the whole day
lcmv_raw for 10min of data processed by chunks takes 1 day???
— Reply to this email directly or view it on GitHub https://github.com/mne-tools/mne-python/issues/2685#issuecomment-163539190 .
Yuval
Dr .Harpaz
BIU MEG lab http://faculty.biu.ac.il/~goldsa/index.html
one option would be to add support for lazy stc as we do for minimum_norm ... it means passing a tuple (kernel, data) to SourceEstimate.
this is done is https://github.com/mne-tools/mne-python/pull/4389
Dear MNE team I am trying to compute LCMV(g2) for epilepsy data. for this I tried using lcmv_raw. I want to compute the beamformer weights on raw data (500sec) and then apply them to short data epochs (0.25sec) , compute g2 (estimator for kurtosis) and save the sum of that as source strength. I can't do it because if I use the whole data I get memory error of course (the stc is too big), and computing LCMV for short data is not the way to go. I suggest doing LCMV in two stages, giving the inv first, then applying it. in the meantime, is there a way for me to get the weights while doing LCMV? thanks