[pymvpa] hyperalignment inquiry

[Swaroop Guntupalli]

Hi David,

If you have limited data, you can use a part of it (however you split the
data for training and testing)
to train hyperalignment, and also use the same part to train the classifier
and then apply hyperalignment and test classifier on the left-out part.
Yes, you can artificially create 2 chunks (or more if you prefer).


On Wed, Jul 27, 2016 at 3:17 PM, David Soto <d.soto.b at gmail.com> wrote:

> sounds great thanks, a further thing is that I have seen that in order to
> preclude  circularity issues, hyperalinment is implemented on a subset of
> training chunks and then the transformation is applied to the full datasets
> prior to classification analyses.  Given that I have no proper chunks/runs
> here, but only 56 betas across trials, would it be okay to train
> hyperaligment just on half of the 56 betas, eg artificially split the data
> set in 2 chunks  each containing 14 betas of class A and 14 of class B? Or
> would it be just OK to train hyperaligment on the 56 betas in the first
> instance?
> thanks!
> david
>
> On 28 July 2016 at 00:00, Swaroop Guntupalli <swaroopgj at gmail.com> wrote:
>
>> The hyperalignment example on PyMVPA uses one beta map for each category
>> per run.
>>
>> On Wed, Jul 27, 2016 at 2:57 PM, Swaroop Guntupalli <swaroopgj at gmail.com>
>> wrote:
>>
>>> Hi David,
>>>
>>> Beta maps should work fine for hyperalignment. The more maps (or TRs)
>>> there are, better the estimate.
>>> We used within-subject hyperalignment in Haxby et al. 2011, which uses
>>> maps from 6 categories (we used 3 successive betas per condition I think).
>>>
>>> vstack() merges multiple datasets into a single dataset, and if there is
>>> any voxel count (nfeatures) mismatch across subjects, it won't work (as
>>> evidenced by the error).
>>> Hyperalignment takes in a list of datasets, one per each subject.
>>> So, you can make that a list as
>>> ds_all =[ds1, ds2, ...., ds16]
>>> and use for Hyperalignment()
>>>
>>> Best,
>>> Swaroop
>>>
>>>
>>> On Wed, Jul 27, 2016 at 2:28 PM, David Soto <d.soto.b at gmail.com> wrote:
>>>
>>>> hi,
>>>>
>>>> in my experiment I have 28 betas in condition A and 28 parameter
>>>> estimate images and 28  in condition B for each subject (N=16 in total).
>>>>
>>>> i have performed across-subjects SVM-based searchlight classification
>>>> using MNI-registered individual beta images and I would like to repeat and
>>>> confirm my results using searchlight based on hyperaligned data.
>>>>
>>>> i am not aware of any paper using hyperaligment on  beta images but I
>>>> think this should be possible, any advise please would be nice
>>>>
>>>> i've created individual datasets concatenating the 28 betas in
>>>> condition A and the 28 in condition (in the actual experiment condition A
>>>> and B can appear randomly on each trial). I have 16 nifti datasets, one per
>>>> subject, with each in individual native anatomical space. In trying to get
>>>> a dataset in the same format as in the hyperlignment tutorial I use
>>>> fmri_dataset on each individual wholebrain 48 betas  and then try to merged
>>>> then all i.e. ds_merged = vstack((d1, d2, d3, d4, d5, d6, d7, d8, d9,
>>>> d10, d11, d12, d13, d14, d15,d16)) but this gives the following error
>>>> pasted at the end,
>>>> which I think it is becos the number of voxels is different across
>>>> subjects. This is one issue.
>>>>
>>>> Another is that the function vstack does appear to produce the list of
>>>> individual datasets that is in the hyperligment tutorial dataset, but a
>>>> list of individual betas, I would be grateful to receive some tips.
>>>>
>>>> thanks!
>>>> david
>>>> ------------------------------------------------------------
>>>> ---------------
>>>> ValueError                                Traceback (most recent call
>>>> last)
>>>> <ipython-input-64-2fef46542bfc> in <module>()
>>>>      19 h5save('/home/dsoto/dsoto/fmri/wmlearning/h5.hdf5', [d1,d2])
>>>>      20 #ds_merged = vstack((d1, d2, d3, d4, d5, d6, d7,d8,d9, d10,
>>>> d11, d12, d13, d14, d15, d16))
>>>> ---> 21 ds_merged = vstack((d1, d2))
>>>>
>>>> /usr/local/lib/python2.7/site-packages/mvpa2/base/dataset.pyc in
>>>> vstack(datasets, a)
>>>>     687                              "datasets have varying
>>>> attributes.")
>>>>     688     # will puke if not equal number of features
>>>> --> 689     stacked_samp = np.concatenate([ds.samples for ds in
>>>> datasets], axis=0)
>>>>     690
>>>>     691     stacked_sa = {}
>>>>
>>>> ValueError: all the input array dimensions except for the concatenation
>>>> axis must match exactly
>>>>
>>>> _______________________________________________
>>>> Pkg-ExpPsy-PyMVPA mailing list
>>>> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
>>>>
>>>> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
>>>>
>>>
>>>
>>
>> _______________________________________________
>> Pkg-ExpPsy-PyMVPA mailing list
>> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
>> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
>>
>
>
> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20160727/4097b7c0/attachment-0001.html>