Adding, removing and moving axes

Reshaping changes how many elements sit on each axis. This lesson is about changing how many axes there are and what order they're in — which sounds like bookkeeping until you meet the two situations where it's the whole job.

Removing length-1 axes

Singleton axes accumulate. a[0:1] leaves one, keepdims=True leaves one, loading a single item out of a batch leaves one.

x = np.zeros((1, 4, 1))
x.squeeze().shape          # (4,)   — drops every length-1 axis
x.squeeze(axis=0).shape    # (4, 1) — drops only that one

Prefer the explicit axis= form in real code. A bare squeeze() silently does nothing when there's nothing to drop, so a shape bug can slide straight past it; naming the axis raises instead.

Adding one

Three spellings, one meaning:

x[:, np.newaxis]           # the readable one
x[:, None]                 # np.newaxis IS None — same thing
np.expand_dims(x, axis=1)  # the function form

The position matters: x[:, None] turns (3,) into (3, 1), while x[None, :] gives (1, 3).

Why any of this matters

Here's the real reason you'll reach for it. Broadcasting lines shapes up from the right. So a (3,) array spreads across each row — and if you wanted it to spread down each column, you have to give it a second axis:

m = np.zeros((3, 3))
m + np.array([1, 2, 3])              # 1,2,3 across every row
m + np.array([1, 2, 3])[:, None]     # 1,2,3 down every column

Same three numbers, perpendicular results. Whenever a broadcast comes out transposed from what you expected, this is why.

Reordering axes

.T reverses all axes. For 2-D that's the transpose you want; for 3-D it turns (2, 3, 4) into (4, 3, 2), which is almost never what anyone means.

np.swapaxes(x, 0, 1)     # exchange two specific axes
np.moveaxis(x, 0, 2)     # take axis 0, put it at position 2
x.transpose(1, 2, 0)     # give the full new order

moveaxis is the one worth remembering, because it reads as a sentence and because of this:

photo.shape                          # (120, 160, 3) — height, width, channels
np.moveaxis(photo, -1, 0).shape      # (3, 120, 160) — channels first

Images arrive channels-last; PyTorch and most convolution code want channels-first. That one line is the conversion, and like every axis operation here it's a view — no data moves.