When two axes share a size
Stack a (512, 512, 3) photo three times and you get a batch of (3, 512, 512, 3). Convert it to NCHW and you get (3, 3, 512, 512).
Two axes now both have size 3 — one counts examples, one counts colours. Once two axes share a size, the shape cannot tell you which is which. Only your own tracking can.
This is the deeper half of the section. Reshape-versus-transpose is a bug you can catch. This one you can only prevent.