Stacking Multiple CNN Blocks

The problem, from first principles

One CNN block (02-one-cnn-block) extracts one level of features and shrinks the spatial size once. A single block rarely learns anything rich enough for a real task — real image classifiers chain several blocks in sequence, each one operating on the PREVIOUS block's output rather than the raw image, so later blocks combine earlier features into progressively higher-level ones (edges → textures → parts → objects) while the spatial size keeps shrinking toward something small enough to flatten.

From theory to code

Theory says: the first block sees the original image, and every later block sees the previous block's output. Chain them, feeding one into the next.

Implement stack_cnn_blocks(image, kernels, pool_size=2) against that reasoning, reusing 02-one-cnn-block's cnn_block once per kernel in the list.

Constraints

• image: shape (C_in, H, W).
• kernels: a list of kernels, one per block. Kernel i's input-channel dimension must match block i's actual input (either image's own channels, for the first block, or the previous block's output channel count).
• pool_size is passed to every block identically.
• An empty kernels list returns image unchanged — zero blocks means no computation at all.
• image and every kernel in kernels are never modified in place.

Hints

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