[Eeglablist] Installing CUDAICA on Windows 10 (2021 update)

Makoto Miyakoshi mmiyakoshi at ucsd.edu
Sat Nov 20 09:56:10 PST 2021


Update--
John Kiat (UC Davis) emailed me who happened to be working on CUDAICA
installation independently in the same period (what a coincidence).
He uploaded a Matlab movie of running CUDAICA to Youtube. If you want to
feel the speed, check it out.
https://urldefense.proofpoint.com/v2/url?u=https-3A__www.youtube.com_watch-3Fv-3DeVgpmvQ9LVU&d=DwIFaQ&c=-35OiAkTchMrZOngvJPOeA&r=kB5f6DjXkuOQpM1bq5OFA9kKiQyNm1p6x6e36h3EglE&m=hNn0oeQyLnH7lF_0qTcOEiiBrn-SwURSUdUvP2HKoG0v_UCZ2wi7HXTDonUb6kg2&s=1KcsNL2LFlPcNNgBNjCVSI6zinY1R1VvAjoE4ylIJs4&e= 
John told me that in his environment (Ryzen 7 3800x, GTX1080ti) he
confirmed x15 boost even after addressing the 'drawnow' slowing issue. That
is a huge difference.
I updated my Wiki section with these link and info.
https://sccn.ucsd.edu/wiki/Makoto%27s_useful_EEGLAB_code#By_using_CUDAICA_.2811.2F20.2F2021_updated:_Thank_you_Ugo.2C_Yunhui.2C_and_John.21.29

Makoto

On Thu, Nov 18, 2021 at 3:48 PM Scott Makeig <smakeig at gmail.com> wrote:

> Makoto and all -
>
> I was writing to the case where the data was to be compressed into a much
> smaller number of dimensions than channels. If the data are rank-deficient,
> reducing the dimension of the data to its true rank using PCA is quite
> acceptable (and necessary to allow ICA decomposition). Sorry if I
> misunderstood the case here. The Artoni paper shows that for the (typical)
> EEG study he was analyzing, removing further dimensions not only reduced
> the number of interpretable ICs but reduced their 'dipolarity', i.e. the
> degree to which they were compatible with a scalp projection from a
> localized cortical source area. I assume this might also reduce their
> independence (i.e., increase the pairwise mutual information of their time
> courses, although I don't think we checked this).
>
> Example 1:  I record 128 channels, then decide that  the signals from 17
> channels are 'bad'. I then decide to linearly interpolate signals for those
> 17. I then convert the recomposed to average reference.  The data rank
> should then be 128-17-1 = 110, since linear interpolation does not add
> *independent information* to the data - and converting to average reference
> loses a further dimension. Applying PCA reduction to 110 dimensions as a
> precursor to ICA is necessary and correct here.
>
> Example 2: I record 128 channels but want to do a 'quick' ICA
> decomposition of dimension 64. So I reduce the data to the largest 64 PCs
> and then perform ICA decomposition. Though the data volume (RMS) accounted
> for by the 64 removed dimensions is very likely small, noticeable
> degradation of the 'brain' ICs results.   This is because the PCs 'cut
> across' all the IC effective sources - and the remaining 64 dimensions sum
> in part the activities of all 128 IC effective source processes, not
> allowing the derived ICs to be truly independent of each other or aligned
> with only one effective source.
>
> Scott
>
> On Thu, Nov 18, 2021 at 6:27 PM Makoto Miyakoshi via eeglablist <
> eeglablist at sccn.ucsd.edu> wrote:
>
>> Dear Scott,
>>
>> Apart from the value of the study, I don't like the side effect Fiorenzo's
>> PCA paper caused: It made non-engineers superstitious about the use of PCA
>> (and now you push this fear campaign.)
>>
>> In the Artoni paper in question, at the very first line of the result
>> section he reported that there were only 8+/-2.5 PCs were left to obtain
>> 95% of variance out of 71 scalp electrodes. You can easily imagine what
>> happens when you reject 63/71 PCs before ICA. In this sense, the
>> conclusion
>> of this study is 'duh' to me (is it not?)
>>
>> 'Applying PCA before ICA is suboptimal' is a qualitative statement. But
>> what if you reject only one or two PCs just to make the data full-ranked?
>> These questions can be answered only by performing numerical and
>> simulation
>> studies. We should educate people to reject qualitative statements and
>> instead think quantitatively.
>>
>> Dear John,
>>
>> It's not about CUDAICA per se, but if you care the obtimality of the ICA
>> results, I recommend you compare results from channel rejection against
>> results from PCA dimension reduction so that you obtain 70 ICs in both
>> results. Compare the results side by side and determine both visually and
>> quantitatively using ICLabel. Most likely, they do not show much
>> difference. Then you feel better to go with PCA because so that you don't
>> need to lose scalp electrodes.
>>
>> Makoto
>>
>>
>>
>> On Wed, Nov 17, 2021 at 1:22 PM Scott Makeig <smakeig at gmail.com> wrote:
>>
>> > John,
>> >
>> > Makoto  seems to forget the result of Fiorenzo Artoni that applying PCA
>> > before ICA is suboptimal - better to reduce the number of channels and
>> run
>> > ICA decomposition full-rank. Or, if you are more ambitious / exacting,
>> run
>> > multiple ICA decompositions on different channel subsets (for example,
>> > random sets of 70 channels picked from 128) and then apply clustering to
>> > the resulting independent component (IC) maps - I haven't seen that
>> > approach applied yet ...
>> >
>> > Artoni, F., Delorme, A. and Makeig, S., 2018. Applying dimension
>> > reduction to EEG data by Principal Component Analysis reduces the
>> quality
>> > of its subsequent Independent Component decomposition
>> > <
>> https://urldefense.proofpoint.com/v2/url?u=https-3A__www.sciencedirect.com_science_article_pii_S1053811918302143-3Fcasa-5Ftoken-3DKT3XImh-2D-2Dl0AAAAA-3Aut0ozF7mVGYDngMVu-2Di0PowjzqzZEhuIl153z6MNgM8NRHDXZZj2CWlYEd0948glBn11q-5FXk7B8&d=DwMFaQ&c=-35OiAkTchMrZOngvJPOeA&r=pyiMpJA6aQ3IKcfd-jIW1kWlr8b1b2ssGmoavJHHJ7Q&m=YlyRvbJOBqbcxuJuFRpFCdKF7Tkg1Qj32rXx8Aa2570eE8IWixleWy79aWfkbQP9&s=n4PV6FRx_PCHi9cIjG7_rLRcb9j62ChP5inSC0zrMnE&e=
>> >
>> > . *NeuroImage*, *175*, pp.176-187.
>> >
>> > Fiorenzo's RELICA plug-in does something related - applying (full-rank)
>> > ICA decomposition to randomly selected subsets of data points, followed
>> by
>> > component clustering. Zeynep Akalin Acar has recently demonstrated that
>> > using RELICA component cluster scalp map means *and* variances  can
>> > increase the accuracy of high-resolution source location estimation.
>> >
>> > Acar, Z.A. and Makeig, S., 2020, October. Improved cortical source
>> > localization of ICA-derived EEG components using a source scalp
>> projection
>> > noise model
>> > <
>> https://urldefense.proofpoint.com/v2/url?u=https-3A__ieeexplore.ieee.org_iel7_9287816_9287978_09288020.pdf-3Fcasa-5Ftoken-3DdOEYtrVfetoAAAAA-3AQsY-2D7AQl9TzSyk3IMqlsy7KnMhUI-2DJ-2DQ68H5cKzjpbPWRVpN-2D4xrR-5FJPMBwDEDF-5FQ9nLkIHLH5U&d=DwMFaQ&c=-35OiAkTchMrZOngvJPOeA&r=pyiMpJA6aQ3IKcfd-jIW1kWlr8b1b2ssGmoavJHHJ7Q&m=YlyRvbJOBqbcxuJuFRpFCdKF7Tkg1Qj32rXx8Aa2570eE8IWixleWy79aWfkbQP9&s=Y8l5g2Z8tUOx5dUtb8pXqMm0HnvqsCYHwlarrNfdP_c&e=
>> >.
>> > In *2020 IEEE 20th International Conference on Bioinformatics and
>> > Bioengineering (BIBE)* (pp. 543-547). IEEE.
>> >
>> > Scott
>> >
>> > On Wed, Nov 17, 2021 at 3:24 PM Makoto Miyakoshi via eeglablist <
>> > eeglablist at sccn.ucsd.edu> wrote:
>> >
>> >> Dear John,
>> >>
>> >> I found it interesting that in your case runica()'s processing time
>> >> linearly increased (63 -> 168 min) as the input data length increased
>> (8
>> >> ->
>> >> 25 min), but that for CUDAICA did not (2.3 -> 2.8 min).
>> >>
>> >> If you have 126 ch, you want to have 126^2*30 = 476280 data points as a
>> >> minimum (from SCCN's never-verified rule of thumb). But you have
>> 275*474=
>> >> 130350 datapoints, which seems suboptimal to ensure a good learning.
>> >> Perhaps you want to apply dimension reduction using PCA to obtain 70
>> ICs,
>> >> so that the same rule of thumb predicts 70^2*30 = 147000 datapoints for
>> >> learning, which is much closer.
>> >>
>> >> Do you want to know more detail about this optimization?
>> >> In fact, without running a simulation you can't theoretically determine
>> >> what number is a good number. This is why I wrote this simulator as an
>> >> EEGLAB plugin. Try it out to 'feel' how much deviation/violation from
>> the
>> >> 'rule of thumb' can negatively impact the decomposition.
>> >>
>> >>
>> https://urldefense.proofpoint.com/v2/url?u=https-3A__www.youtube.com_watch-3Fv-3DCGOw04Ukqws&d=DwIFaQ&c=-35OiAkTchMrZOngvJPOeA&r=kB5f6DjXkuOQpM1bq5OFA9kKiQyNm1p6x6e36h3EglE&m=YfSSJdGbaWUJjsVGL_Bd3kbvoWDALgCGOA44Hn93INujbnQT8WNcijz3CAaY7Km2&s=HuuZdy7O3viOY-fIz_ayjeDkatQ_038Fa2wbfMDeg9I&e=
>> >>
>> >> Makoto
>> >>
>> >> On Tue, Nov 16, 2021 at 2:59 PM Richards, John <
>> RICHARDS at mailbox.sc.edu>
>> >> wrote:
>> >>
>> >> > I was curious about the speed differences for my applications.  I
>> have
>> >> > tested this before, did not write down my results.
>> >> >
>> >> > I ran an EEGlab file, 126 channels * 275 samples * 474 trials, about
>> 8
>> >> min
>> >> > of EEG data.  This is done on a linux node in a a linux cluster,
>> >> Intel(R)
>> >> > Xeon(R) CPU E5-2680 v4 @ 2.40GHz node, with 256G memory, 28 cores.
>> The
>> >> > runica appears to be working on 12 cores.   The gpu was a dual Tesla
>> >> > P100-PCIE-16G.  The cudaica ran on one GPU.
>> >> >
>> >> > runica version took 63 min.  cudaica version took 2 min 15 s;
>> >> > runica appeared to be running on multiple CPUs, ~ 12 CPUs.
>> >> >
>> >> > I concatenated the data for 1422 trials, about 25 min
>> >> > Cudica took 2 min 50s|
>> >> > runica took 2.8 hours, 12 CPUs
>> >> >
>> >> > Most of our runs with infants take 8 to 10 min, some of our adults
>> runs
>> >> > are 25 min.
>> >> >
>> >> > I understand from the earlier conversation the binica might be able
>> to
>> >> > match these results?   I'm not going to do a full test, but this
>> >> convinces
>> >> > me to stick with cudaica for now.
>> >> >
>> >> > John
>> >> >
>> >> >
>> >> > -----Original Message-----
>> >> > From: Richards, John
>> >> > Sent: Thursday, November 11, 2021 1:15 AM
>> >> > To: Makoto Miyakoshi <mmiyakoshi at ucsd.edu>; ugob at siu.edu;
>> >> > eeglablist at sccn.ucsd.edu
>> >> > Subject: RE: [Eeglablist] Installing CUDAICA on Windows 10 (2021
>> update)
>> >> >
>> >> > Re CUDAICA.  I was able to install it, i don't remember it being that
>> >> > difficult.  I had to mess around with the CUDA version.
>> >> >
>> >> > I have found it "blazing" fast compared to runica. I have not timed
>> it.
>> >> > We have 10-15 min sessions with EGI 128, 250 hz, do the Prep
>> pipeline to
>> >> > get avg ref, and then CUDAICA.  It takes < 5 min to do the Prep,
>> and  <
>> >> 5
>> >> > min to do the CUDAICA; cf 45 min to 60 min with runica.  I may not be
>> >> using
>> >> > the most recent runica.   BTW, we have fairly powerful computers; we
>> >> use 48
>> >> > cores for the Prep pipeline which is a vast speedup, and V100's with
>> >> 16gb
>> >> > or 32gb.   Definitely not bargain chips.  We use the 48core computers
>> >> for
>> >> > the runica, but it does not appear to profit from the multiple CPUs.
>> >> The
>> >> > Prep pipeline also is very slow on single CPUs, but very fast on the
>> 48
>> >> CPU
>> >> > machines.
>> >> >
>> >> > I would be glad to share more details if anyone is interested.
>> >> >
>> >> > John
>> >> >
>> >> >
>> >> > ***********************************************
>> >> > John E. Richards
>> >> > Carolina Distinguished Professor
>> >> > Department of Psychology
>> >> > University of South Carolina
>> >> > Columbia, SC  29208
>> >> > Dept Phone: 803 777 2079
>> >> > Fax: 803 777 9558
>> >> > Email: richards-john at sc.edu
>> >> >
>> >> >
>> >>
>> https://urldefense.proofpoint.com/v2/url?u=https-3A__jerlab.sc.edu&d=DwIFAw&c=-35OiAkTchMrZOngvJPOeA&r=pyiMpJA6aQ3IKcfd-jIW1kWlr8b1b2ssGmoavJHHJ7Q&m=XWfhosWnNSjs97eRAV2Ysofk5w2Z2_mbQvfeek3KRqTVlZ-2fBHSCo5P_bnFInes&s=yvIsDcwOpKjhTPokE_cuv5RlAl7bUeNjmpt7-e34zWk&e=
>> >> > *************************************************
>> >> >
>> >> > -----Original Message-----
>> >> > From: eeglablist <eeglablist-bounces at sccn.ucsd.edu> On Behalf Of
>> Makoto
>> >> > Miyakoshi via eeglablist
>> >> > Sent: Thursday, November 11, 2021 1:02 AM
>> >> > To: EEGLAB List <eeglablist at sccn.ucsd.edu>; ugob at siu.edu
>> >> > Subject: [Eeglablist] Installing CUDAICA on Windows 10 (2021 update)
>> >> >
>> >> > Dear list members,
>> >> >
>> >> > I summarized the steps to install cudaica() which uses GPU
>> computation
>> >> to
>> >> > calculate infomax ICA (Raimondo et al., 2012). The result from the
>> speed
>> >> > comparison between runica() and cudaica() was not as dramatic as x25
>> >> > reported by the original paper, probably because Tjerk's smart hack
>> >> alone
>> >> > already gave x4-5 speed up to runica(). Still, using a relatively
>> cheap
>> >> > GTX1660 (the pre-COVID price range is $250), I confirmed x4-5 speed
>> up
>> >> > compared with runica(). The detailed instruction can be found in the
>> >> > following link.
>> >> >
>> >> >
>> >> >
>> >>
>> https://sccn.ucsd.edu/wiki/Makoto%27s_useful_EEGLAB_code#By_using_CUDAICA_.2811.2F10.2F2021_added.29
>> >> >
>> >> > WARNING: The installation was difficult.
>> >> >
>> >> > Makoto
>> >> > _______________________________________________
>> >> > Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>> >> > To unsubscribe, send an empty email to
>> >> > eeglablist-unsubscribe at sccn.ucsd.edu
>> >> > For digest mode, send an email with the subject "set digest mime" to
>> >> > eeglablist-request at sccn.ucsd.edu
>> >> >
>> >> _______________________________________________
>> >> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>> >> To unsubscribe, send an empty email to
>> >> eeglablist-unsubscribe at sccn.ucsd.edu
>> >> For digest mode, send an email with the subject "set digest mime" to
>> >> eeglablist-request at sccn.ucsd.edu
>> >>
>> >
>> >
>> > --
>> > Scott Makeig, Research Scientist and Director, Swartz Center for
>> > Computational Neuroscience, Institute for Neural Computation,
>> University of
>> > California San Diego, La Jolla CA 92093-0559,
>> http://sccn.ucsd.edu/~scott
>> >
>> _______________________________________________
>> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>> To unsubscribe, send an empty email to
>> eeglablist-unsubscribe at sccn.ucsd.edu
>> For digest mode, send an email with the subject "set digest mime" to
>> eeglablist-request at sccn.ucsd.edu
>>
>
>
> --
> Scott Makeig, Research Scientist and Director, Swartz Center for
> Computational Neuroscience, Institute for Neural Computation, University of
> California San Diego, La Jolla CA 92093-0559, http://sccn.ucsd.edu/~scott
>



More information about the eeglablist mailing list