Flatten: Bridging Convolutional and Linear Layers

The problem, from first principles

Everything built in 01-convolutions and 02-pooling operates on spatial data — arrays shaped (C, H, W), where position matters. 01-classical-ml/01-linear-regression/01-hypothesis-function's linear operates on the opposite shape of assumption: a flat vector of features, no spatial structure at all. To connect a convolutional feature extractor to a Linear classifier head, something has to convert one shape convention into the other — that's flatten's entire job.

From theory to code

Theory says there's no real computation here, only a reinterpretation: take every value in a (C, H, W) array and lay it out as one long 1D vector, in the same order the array is already stored in.

Implement flatten(image) against that reasoning.

Constraints

• image: shape (C, H, W).
• Output: shape (C*H*W,).
• Order: channel-major, then row, then column — the same order image.reshape(-1) and torch.flatten both use by default.
• Every value must be preserved exactly, only the shape changes.
• image is never modified in place.

Hints

Open one at a time. Each gives away a little more than the last.