AdaBin: Improving Binary Neural Networks with Adaptive Binary Sets

By Zhijun Tu, Xinghao Chen, Pengju Ren and Yunhe Wang

This is the PyTorch implementation of ECCV 2022 paper "AdaBin: Improving Binary Neural Networks with Adaptive Binary Sets” .

Requirements

torch==1.8.0
torchvision==0.9.0
prefetch_generator
progress

Results

• Classification results on CIFAR-10

Model	Bit-width (W/A)	Accuracy
ResNet-20	1/1	88.1%
ResNet-18	1/1	62.1%
VGG-small	1/1	92.3%

• Classification results on ImageNet-1k (* means using the two-step training setting as ReActNet)

Model	Bit-width (W/A)	Top-1. Acc	Top-5. Acc
AlexNet	1/1	53.9%	77.6%
ResNet-18	1/1	63.1%	84.3%
ResNet-18*	1/1	66.4%	86.5%
ResNet-34	1/1	66.4%	86.6%

Citation

@inproceedings{tu2022adabin,
    title={AdaBin: Improving Binary Neural Networks with Adaptive Binary Sets},
    author={Zhijun Tu, Xinghao Chen, Pengju Ren and Yunhe Wang},
    booktitle={European Conference on Computer Vision (ECCV)},
    year={2022}
}