rsagroup / rsatoolbox Goto Github PK

Should follow after #78

Instead of RDMs this creates DataSet objects. Could be used for various meadows tasks

• detection of file type
• ma trials could be an obs_desc
• scope: multiple tasks
• add descriptors from task
• add descriptors from stimuli
• scope: one participant / multiple (MA) tasks (json only)
• scope: multiple participants / multiple (MA) tasks (json only)

Previous toolboxes set values for comparing stimuli to themselves to nan/ignore instead of 0s as pyrsa does currently. This appears sensible to avoid inflated model performance, as all models will predict 0s for those comparisons.

• should read google style docstrings
• teach it about package structure

Compiling the docs returns this:

pyrsa/docs/source/pyrsa.rst:65: WARNING: toctree contains reference to nonexisting document 'pyrsa.simulation.sim'

• index page
• conf file

Hello,
I cannot access the documentation with the steps provided under "documentation." Is there a way you could upload a PDF or another way to access the documentation?
Thank you,
CJ

implementing a version which calculates RDMs for a list of dataset objects would be nice

For datasets, RDMs and results

can import from Charestlab/pyrsa

should allow:
transparancy & circular aperture
size regulation
colored border
strings as replacements

simple travis configuration to

• run tests
• run linter
• pypi

The RDM pattern_descriptors field seems to be incorrect in its current form. There should be one descriptor per RDM value, and ideally they contain info about the condition pair (e.g. condition 1 vs 2) that the according dissimilarity value in the RDM reflects. However, these are not auto-generated, and instead contain an index of size # conditions.

I don't think the user should be responsible for generating these descriptors. Thoughts?

non-squared RDMs
normalization
from ordinal data

As discussed and seconded in pull request #16 , one design decision is to have a supplementary class called DatasetList object with the base of a list of dataset objects. This include implementation of other principled segmentation operations such as sliding window for dynamic analysis.

• use a config file
• separate docs requirements file, requiring ipykernel and nbsphinx etc

integrates the TODOs from dataset.py

I.e. at least the correlations which take the distribution for 0 rdms into account

We should get the corrections for multiple testing implemented. So far there are none.

The new/old correlation coefficient for rank correlations

Currently the RDMs object supports indexing by integers and bools.

It would be possible to expand this to support indexing by slices (e.g. rdms[2:]), as well as setting by index (e.g. rdms[2:2] = RDMs(…) or something), as well as indexing by descriptors (e.g. rdms['type'], or something pandas-like rdms[rdms['type'] == 'animate']). Also related magic methods such as __delitem__.

Which of these would be useful?

Which of these would violate the principle of least astonishment?

When should we return copies of RDM objects versus references?

I was looking at https://rsa3.readthedocs.io/en/latest/distances.html#crossnobis-dissimilarity , and noticed that the covariance matrix is not included.

Shouldn't the covariance matrix be included in a cross-validated Mahalanobis distance, too? https://en.wikipedia.org/wiki/Mahalanobis_distance

Not sure what this would look like in a cross-validated version.

we don't need a dummy test anymore as we now have several regular tests

The current version refers to a property called measurements which does not seem to be in use.

And when this is removed it is unclear how the following line:

dissimilarities[:, order][:, :, order]

would sort the pattern order

https://codeclimate.com/repos/5d9b9be58feef31572014b94/settings/exclude_patterns/edit

At this point, the automatic unit tests, codeclimate and lint style checks by GitHub doesn't seem to be working any more. Any experts looking into this would be appreciated.

It would be great to have them as a coding guideline for contributors, especially because eventually we will officially launch this software package to the general public.

To integrate the weighted MDS into the sklearn module, instead of maintaining it in pyRSA.
Reference PR#97: #97

Create a trivial test so we can test the test commands :p

Perhaps use math directive:

http://docutils.sourceforge.net/docs/ref/rst/directives.html#math

The code should be:

def get_unique_unsorted(array):
    """return a unique unsorted list
    """
    u, indices = np.unique(array, return_index=True)
    temp = indices.argsort()
    return u[temp]

Otherwise the initial order of observation descriptors does not get restored.

Perhaps something to do with imports on rtd

At least:

• Euclidean distance
• Correlation distance
• Mahalanobis distance
• Crossnobis distance

At least:

• correlation distance
• rank correlation distance
• Kendal-Tau
• cosine distance

when dealing with a large number of conditions, e.g. with the natural scenes dataset, indicator will fail with a memory issue.

should we have a fallback plan to compute diff_1, and diff_2 the slow (but memory friendly) way if measurements1.shape[0] > something ?

c_matrix = allpairs(np.arange(measurements1.shape[0]))
diff_1 = np.matmul(c_matrix, measurements1)
diff_2 = np.matmul(c_matrix, measurements2)

My understanding is that the index_vector passed to any function in indicator.py is a column vector (of same size as the number of rows in data.measurements), but it seems the indicator functions only work with row vectors.
Given my limited python experience, I don't see a fully satisfactory fix to allow dynamic handling of row or column vectors. Any insights? Or should we always assume a specific dimensionality?

The first implementations of the different rdm calculations largely assume that the rdms all have the same objects in them. We should change that if we want to be stable to unbalanced designs later

Hi @HeikoSchuett ,

given that we use einsum for all other RDM calc cases, I figured we could also use it for correlation distance. that speeds up performance massively compared to numpy corrcoef.

def correlation_distance(X):
    """[correlation distance]

    Args:
        X ([dataset]): conditions x voxels dataset

    Returns:
        RDM in upper triangular form.        
    """
    X = X - X.mean(axis=1, keepdims=True)
    X /= np.sqrt(np.einsum('ij,ij->i', X, X))[:, None]

    m = X.shape[0]
    r, c = np.triu_indices(m, 1)
     
    return 1 - np.einsum('ik,jk', X, X))[r, c]

i tried to save a results object using its save method and it didn't immediately work.
i haven't investigated. here's the error and traceback...

File "C:\Users\Niko Kriegeskorte\Google Drive\programming\pyRSA\pyrsa\demos\exercise_all.py", line 311, in
results_3_full.save('results_alexnetSim2')

File "c:\users\niko kriegeskorte\google drive\programming\pyrsa\pyrsa\pyrsa\inference\result.py", line 67, in save
write_dict_hdf5(filename, result_dict)

File "c:\users\niko kriegeskorte\google drive\programming\pyrsa\pyrsa\pyrsa\util\file_io.py", line 25, in write_dict_hdf5
_write_to_group(file, dictionary)

File "c:\users\niko kriegeskorte\google drive\programming\pyrsa\pyrsa\pyrsa\util\file_io.py", line 41, in _write_to_group
_write_to_group(subgroup, value)

File "c:\users\niko kriegeskorte\google drive\programming\pyrsa\pyrsa\pyrsa\util\file_io.py", line 41, in _write_to_group
_write_to_group(subgroup, value)

File "c:\users\niko kriegeskorte\google drive\programming\pyrsa\pyrsa\pyrsa\util\file_io.py", line 41, in _write_to_group
_write_to_group(subgroup, value)

File "c:\users\niko kriegeskorte\google drive\programming\pyrsa\pyrsa\pyrsa\util\file_io.py", line 41, in _write_to_group
_write_to_group(subgroup, value)

File "c:\users\niko kriegeskorte\google drive\programming\pyrsa\pyrsa\pyrsa\util\file_io.py", line 38, in _write_to_group
group[key] = value

File "C:\Users\Niko Kriegeskorte\Anaconda3\envs\PyRSA\lib\site-packages\h5py-2.10.0-py3.7-win-amd64.egg\h5py_hl\group.py", line 387, in setitem
ds = self.create_dataset(None, data=obj, dtype=base.guess_dtype(obj))

File "C:\Users\Niko Kriegeskorte\Anaconda3\envs\PyRSA\lib\site-packages\h5py-2.10.0-py3.7-win-amd64.egg\h5py_hl\group.py", line 136, in create_dataset
dsid = dataset.make_new_dset(self, shape, dtype, data, **kwds)

File "C:\Users\Niko Kriegeskorte\Anaconda3\envs\PyRSA\lib\site-packages\h5py-2.10.0-py3.7-win-amd64.egg\h5py_hl\dataset.py", line 118, in make_new_dset
tid = h5t.py_create(dtype, logical=1)

File "h5py\h5t.pyx", line 1634, in h5py.h5t.py_create

File "h5py\h5t.pyx", line 1656, in h5py.h5t.py_create

File "h5py\h5t.pyx", line 1717, in h5py.h5t.py_create

TypeError: No conversion path for dtype: dtype('<U5')

Hi,

In order to define our data set, could you please provide us with a step-by-step for how to import fMRI data an put it in the right format for the toolbox?

https://github.com/rsagroup/pyrsa/community

We should implement Kendall-tau A

this path generated by sphinx is currently show, instead we are ignoring docs/source/build