[pymvpa] hyperalignment inquiry
David Soto
d.soto.b at gmail.com
Thu Jul 28 17:51:43 UTC 2016
Thanks Swaroop, I managed to get the dataset in the right format as per the
hyperaligmentsearchlight tutorial
however when I run the hyperaligment I get the following error (IndexError:
index 46268 is out of bounds for axis 1 with size 43506, see further
below)...to recap the dataset is a concatenation of each subject data, each
in individual native space, so number of features are different across
subjects...
The code I use is the same as in the tutorial, namely, any feedback would
be great, thanks, david
cv = CrossValidation(clf, NFoldPartitioner(attr='subject'),
errorfx=mean_match_accuracy)
for test_run in range(nruns):
ds_train = [sd[sd.sa.chunks != test_run, :] for sd in ds_all]
ds_test = [sd[sd.sa.chunks == test_run, :] for sd in ds_all]
slhyper = SearchlightHyperalignment(radius=3, featsel=0.4,
sparse_radius=3)
slhypmaps = slhyper(ds_train)
ds_hyper = [h.forward(sd) for h, sd in zip(slhypmaps, ds_test)]
ds_hyper = vstack(ds_hyper)
zscore(ds_hyper, chunks_attr='subject')
res_cv = cv(ds_hyper)
bsc_slhyper_results.append(res_cv)
OUTPUT MESSAGE.........
Performing classification analyses...
between-subject (searchlight hyperaligned)...
---------------------------------------------------------------------------
IndexError Traceback (most recent call last)
<ipython-input-191-85bdb873d4f1> in <module>()
24 # Searchlight Hyperalignment returns a list of mappers
corresponding to
25 # subjects in the same order as the list of datasets we passed
in.
---> 26 slhypmaps = slhyper(ds_train)
27
28 # Applying hyperalignment parameters is similar to applying any
mapper in
/usr/local/lib/python2.7/site-packages/mvpa2/algorithms/searchlight_hyperalignment.pyc
in __call__(self, datasets)
626 node_blocks = np.array_split(roi_ids, params.nblocks)
627 p_results = [self._proc_block(block, datasets,
hmeasure, queryengines)
--> 628 for block in node_blocks]
629 results_ds = self.__handle_all_results(p_results)
630 # Dummy iterator for, you know, iteration
/usr/local/lib/python2.7/site-packages/mvpa2/algorithms/searchlight_hyperalignment.pyc
in _proc_block(self, block, datasets, featselhyper, queryengines, seed,
iblock)
387 continue
388 # selecting neighborhood for all subject for
hyperalignment
--> 389 ds_temp = [sd[:, ids] for sd, ids in zip(datasets,
roi_feature_ids_all)]
390 if self.force_roi_seed:
391 roi_seed =
np.array(roi_feature_ids_all[self.params.ref_ds])
== node_id
/usr/local/lib/python2.7/site-packages/mvpa2/datasets/base.pyc in
__getitem__(self, args)
139
140 # let the base do the work
--> 141 ds = super(Dataset, self).__getitem__(args)
142
143 # and adjusting the mapper (if any)
/usr/local/lib/python2.7/site-packages/mvpa2/base/dataset.pyc in
__getitem__(self, args)
445 if isinstance(self.samples, np.ndarray):
446 if np.any([isinstance(a, slice) for a in args]):
--> 447 samples = self.samples[args[0], args[1]]
448 else:
449 # works even with bool masks (although without
IndexError: index 46268 is out of bounds for axis 1 with size 43506
On 28 July 2016 at 00:25, Swaroop Guntupalli <swaroopgj at gmail.com> wrote:
> 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
>>
>
>
> _______________________________________________
> 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/20160728/7d501873/attachment-0001.html>
More information about the Pkg-ExpPsy-PyMVPA
mailing list