Parameter Counting

The problem, from first principles

"How big is this model?" is the very first question anyone asks about a network before they even train it — does it fit in memory, is it comparable to a known baseline, how much would fine-tuning it cost. The answer is a single number: the total count of individual learnable scalar values across every weight and bias tensor.

From theory to code

Implement count_parameters(params), the total element count across a list of parameter arrays, and count_trainable_parameters(params, requires_grad), the same count restricted to parameters that are actually trainable (not frozen).

Constraints

• params: a list of NumPy arrays of any shapes.
• count_parameters returns the sum of every array's total element count.
• requires_grad: a list of booleans, one per entry in params.
• count_trainable_parameters returns the sum of element counts only where the matching requires_grad entry is True.
• An empty params list returns 0.

Hints

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