IC-627ResNet 18 and ViT B/16 retain significant ImageNet accuracy at 2–3 bit weight compression via JLCM

Edouard YVINEC, Arnaud Dapogny, Kevin Bailly

SourceNetwork Memory Footprint Compression Through Jointly Learnable Codebooks and Mappings

The paper measures ImageNet top-1 accuracy of ResNet 18 and ViT B/16 after compressing their weights with JLCM at two compression rates. At 5.33x compression (approximately 3 bits), ResNet 18 achieves 62.939% and ViT B/16 achieves 80.558%. At 7.5x compression (approximately 2 bits), ResNet 18 drops to 41.779% and ViT B/16 to 69.996%. The authors report that JLCM reaches over 99% of the original ViT B/16 accuracy in its data-driven variant at 5.33x.

Evidence
correlational
Key metric
ResNet 18: 62.939% at 5.33x, 41.779% at 7.5x; ViT B/16: 80.558% at 5.33x, 69.996% at 7.5x
Caveat
Results depend on the calibration set size and the specific JLCM hyperparameters; the data-free initialization alone performs substantially worse than the full data-driven JLCM.
Model
ResNet / ResNet-152 / ResNet-101 / ResNet-50-BN ResNet-18, ViT ViT-B/16
Datasets
ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval]
Methods
NUPES [compared-to], BRECQ [compared-to], ADRound [compared-to], PowerQuant [compared-to], RED++ [compared-to], SQuant [compared-to], DFQ [compared-to]
Related work
NUPES [compared-to], BRECQ [compared-to], PowerQuant [compared-to], RED++ [compared-to]
Related findings
IC-626
Extraction
automatic-extraction