Unable to compute the cosine with stats.compare_images

[me-pic]

Issue summary

Error encountered while trying to use the cosine to compute stats.compare_images.

Detailed issue description

Hello Neuromaps team,

I encountered an error while I was trying to reproduce the tutorial Comparing brain maps specifying cosine as the chosen metric (see also steps below).

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[16], line 1
----> 1 cossim = 1 - compare_images(ns_res, aba_res, metric=cosine)

File ~/Documents/master_analysis/brainmaps/lib/python3.9/site-packages/neuromaps/stats.py:61, in compare_images(src, trg, metric, ignore_zero, nulls, nan_policy)
     59     else:
     60         if not isinstance(metric([1, 1], [1, 1]), float):
---> 61             raise ValueError('Provided callable `metric` must accept two '
     62                              'inputs and return single value.')
     64 srcdata, trgdata = load_data(src), load_data(trg)
     66 mask = np.zeros(len(srcdata), dtype=bool)

ValueError: Provided callable `metric` must accept two inputs and return single value.

The problem seems to be with the verification of the variable method in stats.py (L60). Indeed, if I run cosine([1, 1], [1, 1]) I obtain 0 (int), which lead to not isinstance(cosine([1, 1], [1, 1]), float) equally True, thus raising the ValueError. For the moment I just bypass the problem by changing L60 for if not isinstance(metric([1, 2], [1, 1]), float):.

Don't know if this info is relevant, but I'm using scipy==1.10.1

Thank you for your help and for this really cool tool !

Steps to reproduce issue

>>> from neuromaps.datasets import fetch_annotation
>>> neurosynth = fetch_annotation(source='neurosynth')
>>> abagen = fetch_annotation(source='abagen')
>>> from neuromaps.resampling import resample_images
>>> ns_res, aba_res = resample_images(src=neurosynth, trg=abagen,
...                                   src_space='MNI152', trg_space='fsaverage',
...                                   method='linear', resampling='downsample_only')
>>> from scipy.spatial.distance import cosine
>>> cossim = 1 - compare_images(ns_res, aba_res, metric=cosine)

Software version

0.0.3

What operating system were you using when you encountered this issue?

• macOS
• Windows
• Linux

Code of Conduct

• I agree to follow the neuromaps Code of Conduct

[zhangjb35]

It is simple. Just change one number. change stats.py:

isinstance(metric([1, 1], [0, 1]), float)

[justinehansen]

Sorry it took me so long to get to this 🥲
The issue is that neuromaps.stats.compare_images is double checking that the metric you use (if it isn't pearsonr or spearmanr) does accept two array inputs and returns a single float. To test this we use two dummy arrays which are [1, 1] and [1, 1]. This issue surfaces because cosine([1, 1], [1, 1]) is 0 (int) so it fails the condition. So as a fix I am changing it to double check that the output of the metric is a scalar (instead of a float). (The proposed solution above would also work.)

(scipy.spatial.distance.cosine should probably be fixed such that the output of cosine([1, 1], [1, 1]) is 0.0 instead of 0, to be consistent).

Hope this is helpful even after many months.

Justine