Edge models — magnitude pruning

Background

To run on phones and wearables, vision models are often pruned: small-magnitude weights contribute little, so zeroing them shrinks the model with minimal accuracy loss. The simplest rule is magnitude pruning — drop any weight whose absolute value is below a threshold.

Problem statement

Implement magnitude_prune(W, threshold) returning a copy of W with every entry whose absolute value is < threshold set to 0.

Input

• W — array-like weight tensor.
• threshold — float $\ge 0$.

Output

Returns an np.ndarray the same shape as W: weights with $|w| \ge \text{threshold}$ kept, the rest zeroed.

Examples

Example 1

Input:  W = [[0.5, -0.05], [0.2, -0.8]], threshold = 0.1
Output: [[ 0.5  0. ], [ 0.2 -0.8]]

Explanation: $|-0.05| < 0.1$ is pruned to 0; the rest survive.

Example 2 — threshold 0 keeps everything

Input:  W = [1, -2, 3], threshold = 0
Output: [1. -2.  3.]

Constraints

• Keep entries with $|w| \ge \text{threshold}$; zero the rest.
• Return an np.ndarray of the same shape.

Notes

• Pruning is one of the standard edge-optimization levers, alongside quantization (lower precision) and distillation (small student mimics a big teacher).