keep_nans=False ignored in nearest_neighbor mode

[sametSeckiiin]

Hi,

I noticed that SequenceInterpolator doesn't actually handle NaNs when using nearest_neighbor mode, even if keep_nans=False is set.

In interpolators.py, the linear mode has a check to fill NaNs using nanmean, but the nearest_neighbor block just returns the raw data from idx_lower:
if self.interpolation_mode == "nearest_neighbor":
data = self._data[idx_lower]
return (data, valid) if return_valid else data

because of this, interpolate() still returns NaNs when it's supposed to fill them. Is this a bug or am I missing something?

Also, idx_lower uses np.floor, so it’s always looking at the previous index instead of the actual "nearest" one. Just wanted to point that out.

[pollytur]

Hi @sametSeckiiin and thanks for your interest in experanto!
Please give us a link or a minimal example to reproduce the issue

[sametSeckiiin]

Sure, here is a minimal snippet to reproduce it:

import numpy as np
from experanto.interpolators import SequenceInterpolator

dummy data with a NaN

t = np.array([0.0, 1.0, 2.0])
data = np.array([10.0, np.nan, 30.0])

interp = SequenceInterpolator(t, data, interpolation_mode="nearest_neighbor", keep_nans=False)

interpolating near t=1.0
print(interp.interpolate(np.array([1.2])))

If you run this, it outputs [nan]. Since keep_nans=False, I'd expect it to handle the NaN (like how linear mode uses nanmean) but it directly pulls the raw NaN from the data array.

Let me know if this makes sense. I'd be happy to open a PR to fix this

[pollytur]

@binary69 could you please take a look at this somewhere this week?

[binary69]

@pollytur it's mention in the doc string "For 'nearest_neighbor', NaNs are left unchanged."

About the second problem yup having np.round instead of np.floor would work as for example, if it's to choose nearest for 0.7, nearest between (0.0,1.0) it'd give 0.0 instead of 1.0 with the current floor

[pollytur]

ok, thanks!
for the floor - we probably keep it as it is for now (might come back to it later)
for knn interpolation and keep_nans we probably want to add a warning then, as it is an unexpected behaviour

[sametSeckiiin]

Thanks for the explanation

[pollytur]

@binary69 could you please make the PR with the warning? no rush, you can do it after the IBL export draft PR