where, select and clip

A mask selects and loses the shape. Very often you want the opposite: keep every position and replace the values that don't qualify. That's an if-statement applied to a whole array, and np.where is how you write one.

where

np.where(condition, value_if_true, value_if_false)

It walks every position and picks from the second or third argument. The result has the same shape as the condition:

np.where(scores >= 60, 1, 0)             # a pass/fail grid
np.where(scores >= 60, scores, 60)      # lift failures up to 60
np.where(scores >= 60, scores, 0)       # zero out failures

All three arguments broadcast, so you can mix scalars and arrays freely — including two different arrays, which is how you'd merge two datasets by a condition.

Compare the two tools directly:

scores[scores >= 60]              # shape (3,)   — selects, flattens
np.where(scores >= 60, 1, 0)      # shape (2, 3) — replaces, keeps shape

where with one argument

Called with only a condition, np.where does something different: it returns the coordinates of the True positions, as one array per axis.

np.where(scores >= 80)     # (array([0, 1]), array([2, 2]))
np.nonzero(scores >= 80)   # identical, and better named

Those tuples are ready to use as fancy indices, which ties lesson 8 to this one. Prefer np.nonzero when that's what you mean — the one-argument where is a historical accident and reads confusingly next to the three-argument form.

clip

Squeezing values into a range is common enough to have its own function, and it says what it means:

scores.clip(40, 80)                        # both ends
scores.clip(min=40)                        # floor only
np.clip(photo.astype(int) + 60, 0, 255)    # the image-brightening pattern

More than two outcomes

Nesting np.where inside itself gets unreadable after the second level. For several bands, use np.select — a list of conditions and a matching list of results, evaluated in order, first match wins:

np.select(
    [scores >= 80, scores >= 60, scores >= 40],
    [3,            2,            1           ],
    default=0,
)

Order matters. Because the first match wins, put the most specific condition first — with these bands reversed, everything above 40 would score 1.

Writing through a mask

You can also assign directly into the matching positions:

scores[scores < 40] = 40         # in place
np.putmask(scores, scores < 40, 40)   # the same thing, spelled out

The difference from np.where matters: np.where returns a new array and leaves the original alone, while assignment edits in place — and therefore edits every view onto that memory, exactly as in lesson 6. Reach for np.where by default, and assign in place only when you mean to.