CNN compression for edge deployment

why i built it

This was a college project measuring how model compression techniques affect accuracy, file size, and CPU inference speed.

what broke

The first challenge was fitting the baseline training run within Colab's compute limits. Runs at 200 and then 175 epochs crashed; 150 epochs completed.

benchmarks

• Baseline ResNet-18

  95.34% top-1 accuracy · 44.77 MB · 28.2 ms CPU latency

• 70% unstructured pruning

  94.94% accuracy, with no measurable model-size or latency reduction

• 50% structured channel pruning

  94.25% accuracy · 10.73 MB · 1.62× speedup after fine-tuning

• Structured pruning + INT8 quantization

  94.32% accuracy · 2.76 MB · 2.07× speedup · 15.5× smaller than baseline

• Knowledge distillation

  Did not improve on the fine-tuned structured-pruning model