Boolean masks

Comparing an array to a number compares every element and hands back an array of the same shape full of True and False. That's a boolean mask, and its dtype is bool — one byte per element.

Put the mask in the brackets and you get the elements sitting under a True.

The result is 1-D

cups[cups > 100] on a (4, 3) array gives shape (5,) — not a smaller rectangle. It has to. The surviving cells are scattered; they don't form a rectangle, and unlike pandas there's no index to record where the holes were. So NumPy flattens.

If you want the shape preserved, don't select — replace. That's np.where, and it gets the lesson after next.

Combining conditions

Two traps, both in the same line of code. Use &, | and ~ — not and, or, not, which demand a single true-or-false answer and can't get one from an array. And parenthesise each condition, because & binds tighter than >:

cups[(cups > 90) & (cups < 130)]      # correct
cups[cups > 90 & cups < 130]          # ValueError

For a list of allowed values, np.isin beats a chain of |:

cups[np.isin(cups, [60, 90, 120])]

Masks are numbers too

True is 1, so a mask supports arithmetic — and that gives you counting for free:

(cups > 100).sum()            # how many match
(cups > 100).sum(axis=0)      # how many per column
(cups > 100).any()            # at least one?
(cups > 100).all()            # every one?
np.count_nonzero(cups > 100)  # the same count, stated plainly

.any() and .all() take an axis too, which is often exactly what you want: "which rows have any missing value?" is np.isnan(x).any(axis=1).

Masks select rows, as well as elements

A 1-D mask whose length matches an axis selects along that axis, and this form keeps the rectangle:

busy = cups.sum(axis=1) > 300   # one True/False per day, shape (4,)
cups[busy]                      # the matching rows — still 2-D

That's the pattern you'll use constantly on real data: build a per-row condition, then use it to keep rows. It stays rectangular because you're selecting whole rows, not scattered cells.

Finally, masks work on the left of an assignment too — cups[cups < 70] = 70 edits in place. Powerful, and it changes the original array along with every view of it.