[pymvpa] hyperalignment inquiry
Swaroop Guntupalli
swaroopgj at gmail.com
Fri Jul 29 14:17:57 UTC 2016
Hi David,
If you are using searchlight hyperalignment, it is advisable to align the
data across subjects using anatomy first. Simplest would be to be align
them to an MNI template and then run the searchlight hyperalignment.
Our tutorial dataset is affine aligned to MNI template.
Best,
Swaroop
On Thu, Jul 28, 2016 at 10:51 AM, David Soto <d.soto.b at gmail.com> wrote:
> 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
>>
>
>
> _______________________________________________
> 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/20160729/264860b6/attachment-0001.html>
More information about the Pkg-ExpPsy-PyMVPA
mailing list