Awesome

TBN

This is the implementation of TBN/TBNv2.

Requirements

Install PyTorch and torchvision
Download the ImageNet dataset from http://www.image-net.org/
- Then, and move validation images to labeled subfolders, using the following shell script

Results

Arch	top-1 accuracy	top-5 accuracy
AlexNet (full)	61.6	82.9
AlexNet (TBNv2)	54.9	77.8
PreActResNet18 (full)	70.3	89.3
PreActResNet18 (TBNv2)	59.7	82.1
PreActResNet34 (full)	73.3	76.4
PreActResNet34 (TBNv2)	63.4	849
PreActResNet50 (full)	76.4	93.2
PreActResNet50 (TBNv2)	66.6	86.7

Evaluate

Download pretrained model
- alexnet_TBNv2: BaiduPan (password: 3rh2)
- preact_resnet_18_TBNv2: Coming soon
- preact_resnet_34_TBNv2: Coming soon
- preact_resnet_50_TBNv2: BaiduPan (password: cmq7)
run command (see scripts/*)

python3 ./imagenet.py --load pretrained/alexnet_TBNv2.pth -e -a alexnet --gpu 0 ~/data/ImageNet

Train

Please refer pytorch/examples/imagenet

Use --ternary-delta=0.5, --ternary-order=2, --ternary-momentum=0.1 and --ternary-no-scale to set the hyper-parameter of TBN/TBNv2.

For TBN, you need --ternary-no-scale option.
when --ternary-momentum <= 0, the threshold value of ternary is fixed as --ternary-delta rather than calculating based on inputs.

Citation

@InProceedings{Wan_2018_ECCV,
author = {Wan, Diwen and Shen, Fumin and Liu, Li and Zhu, Fan and Qin, Jie and Shao, Ling and Tao Shen, Heng},
title = {TBN: Convolutional Neural Network with Ternary Inputs and Binary Weights},
booktitle = {The European Conference on Computer Vision (ECCV)},
month = {September},
year = {2018}
}