An Issue About ind2_ Processing in Function computing_indicator()

[zyyk78]

Function computing_indicator() in caiman/source_extraction/cnmf/spatial.py
ind2_ forget to remove background components with A_in.dtype == bool in line 1030

1030          ind2_ = [np.hstack((np.where(iid_)[0], nr + np.arange(f.shape[0])))      
1031                   if np.size(np.where(iid_)[0]) > 0 else [] for iid_ in dist_indicator]   

here ind2_ didn't remove iid>=nr , but

1047        ind2_ = [np.where(iid_.squeeze())[0]  for iid_ in dist_indicator.astype(bool).toarray()]         
1048        ind2_ = [iid_ if (np.size(iid_) > 0) and (np.min(iid_) < nr) else [] for iid_ in ind2_]  

here removed them while A_in is not bool
this may cause some bugs in function update_spatial_components() while using seeded CNMFE (with imported Ain)
in line 163 ind2_ = [ind_list[np.setdiff1d(a,ff)] if len(a) else a for a in ind2_]

I added ind2_ = [iid_[iid_ < nr] if len(iid_) > 0 else [] for iid_ in ind2_] after line 1031 for fixing it

[ethanbb]

See also #1422 - I think this fixes a separate but related issue.

[ethanbb]

Looking more closely - your proposed change would eliminate the need for #1422 and is probably more correct. I wanted to avoid changing behavior as much as possible, but I noted in that PR:

However, I'm not 100% sure that the behavior of adding indices for background components is correct in the first place, since it seems inconsistent with what happens with no masks passed in.

[zyyk78]

ok, seem to be this bug.
But dose it work? It seems work differently with Ain.dtype != bool (im
not sure)

[ethanbb]

I stared at it a bit more. I'm not sure exactly what the computing_indicator function does for non-bool Ain (I only use boolean masks), but it shouldn't matter - the change I made shouldn't change behavior if ind2_ doesn't contain indices >= nr and will prevent an error if it does.

i.e. let's say we have 10 components & 2 background components; we know we get back an ind2_ which is a list of lists of component indices.

Let's say component 1 has a nan, so we want to eliminate it. The goal is to remove any 1s from ind2_, then shift the rest back by subtracting 1 from each component index in ind2_ that is > 1. The line that does this is (slightly simplified): ind2_ = [ind_list[np.setdiff(a,ff)] for a in ind2_] (where ff = [1]). Our ind_list previously would be:

[0, 0, 1, 2, 3, 4, 5, 6, 7, 8]

The change avoids out-of-bounds errors in this line by expanding it to include background components, making it:

[0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]

If ind2_ only contains indices from 0 to nr-1, to begin with, this doesn't change anything.

That being said, I am still unsure whether it is intended to include these background components for the boolean case.

[ethanbb]

See further discussion on the PR

[zyyk78]

Thank you very much! I see. I noticed that while processing non-bool A_in in computing_indicator()
ind2_ = [iid_ if (np.size(iid_) > 0) and (np.min(iid_) < nr) else [] for iid_ in ind2_]
just removed components completely without iid_<nr , but not those idx > nr in each component.
So this may contain background components. (I didn't notice this before)
This means your code is likely correct! And i may have a try

[zyyk78]

I mean, its raw code processing non-bool works like this:
nr = 5
[5] will be removed
[2,5] won't change

But my code works like:
[5] --> []
[2,5] --> [2]
This seems incorrect

[zyyk78]

It seems ok now