NaN, inf and missing values

np.nan — "not a number" — is how a float array says I don't know. A sensor dropped a reading, a field was blank, a division didn't work.

It's a float, always. An integer array cannot hold one, which is why loading a CSV with gaps hands you float columns whether you wanted them or not.

It spreads, and that's correct

x = np.array([3.0, np.nan, 7.5])
x.sum()      # nan
x.mean()     # nan
x.max()      # nan

Not a bug. If one value is genuinely unknown then the true total is genuinely unknown, and NumPy refuses to guess. The alternative — silently pretending the gap is a zero — would give you a confidently wrong answer, which is worse.

The equality trap

Two unknowns aren't equal to each other, so:

np.nan == np.nan    # False
x == np.nan         # all False — finds nothing, raises nothing

That's the dangerous one: you write a filter, it silently matches zero rows, and nothing tells you. Use the function that asks the right question:

np.isnan(x)              # True where the gaps are
x[~np.isnan(x)]          # keep only the real values
np.isnan(x).sum()        # how many are missing
np.isnan(m).any(axis=1)  # which rows have any gap at all

The nan-functions

Rather than masking every time, most reductions have a gap-skipping twin:

np.nansum(x)     np.nanmean(x)    np.nanstd(x)
np.nanmax(x)     np.nanmin(x)     np.nanmedian(x)

np.nanmean divides by the number of real values, so it matches what you'd get by dropping the gaps first. Worth being deliberate about that: skipping a missing value and treating it as zero are different claims about your data, and only one of them is usually true.

Filling gaps

x[np.isnan(x)] = 0                  # explicit
np.nan_to_num(x)                    # nan->0, inf->a huge finite number
np.nan_to_num(x, nan=np.nanmean(x)) # fill with the mean instead

inf is a different thing

np.inf isn't unknown — it's unbounded. Dividing by zero in an array produces it with a RuntimeWarning rather than raising:

np.array([1.0]) / 0    # [inf]  — with a warning
np.array([0.0]) / 0    # [nan]  — 0/0 really is undefined

Before feeding data to anything that will choke on it, the check you usually want catches both problems at once:

np.isfinite(x).all()      # True only if no nan and no inf
x[np.isfinite(x)]         # keep only usable values