Structured Pruning AI Prompt

Apply structured pruning to reduce this model's size and inference cost.

Unstructured pruning (individual weight zeroing) does not improve real-world latency without sparse hardware. Structured pruning removes entire filters, channels, or layers for actual speedup.

1. Sensitivity analysis:
   - For each layer, measure accuracy impact of removing that layer entirely
   - Rank layers from least to most sensitive
   - Layers early in the network and the last classification layer are typically most sensitive

2. Filter/channel pruning (for CNNs):
   - Score filters by L1-norm of weights (smaller norm = less important)
   - Remove the bottom {{prune_ratio}}% of filters in each prunable layer
   - Handle channel dimension changes: update the next layer's input channels accordingly
   - Re-run BN calibration after pruning

3. Attention head pruning (for transformers):
   - Score attention heads by mean attention entropy or gradient-based importance
   - Remove the {{num_heads_to_prune}} least important heads per layer
   - Adjust head projection dimensions accordingly

4. Iterative pruning and fine-tuning:
   - Prune → fine-tune → prune → fine-tune (gradual pruning is better than one-shot)
   - Use a cosine pruning schedule that increases sparsity gradually
   - Target sparsity: {{target_sparsity}}%

5. Results measurement:
   - FLOPs reduction
   - Parameter reduction
   - Inference latency reduction (must measure on real hardware, not estimate)
   - Accuracy change vs unpruned model

Return: sensitivity analysis, pruning implementation, fine-tuning loop, and results table.