CPUBone

ImageNet-1k classification weights for the CPUBone model family (CVPR 2025), converted to the timm implementation of the architecture.

All checkpoints here are the final EMA weights (decay 0.9998) from the original training runs, converted to safetensors in the timm state-dict layout.

Variants

Weights	Params	ImageNet top-1	ImageNet top-5
cpubone_nano.safetensors	6.5M	72.80	90.63
cpubone_b0.safetensors	10.4M	77.40	93.57
cpubone_b1.safetensors	12.4M	78.54	94.05
cpubone_b2.safetensors	30.4M	81.20	95.45
cpubone_b3.safetensors	40.7M	83.03	96.37

Accuracies were measured on the ImageNet-1k validation split with the timm implementation and its default eval preprocessing for these weights (224x224, bicubic, crop_pct 0.95), and match the original training logs within ±0.16pp.

Usage

Requires a timm version that includes the cpubone models:

import timm

model = timm.create_model("cpubone_b0", pretrained=True).eval()