Skip to content

songhan/SqueezeNet-Residual

Repository files navigation

SqueezeNet-Residual

The repo contains the residual-SqueezeNet, which is obtained by adding bypass layer to SqueezeNet_v1.0. Residual-SqueezeNet improves the top-1 accuracy of SqueezeNet by 2.9% on ImageNet without changing the model size(only 4.8MB).

Related repo and paper

SqueezeNet

SqueezeNet-Deep-Compression

SqueezeNet-Generator

SqueezeNet-DSD-Training

SqueezeNet-Residual

If you find residual-SqueezeNet useful in your research, please consider citing the paper:

@article{SqueezeNet,
  title={SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5MB model size},
  author={Iandola, Forrest N and Han, Song and Moskewicz, Matthew W and Ashraf, Khalid and Dally, William J and Keutzer, Kurt},
  journal={arXiv preprint arXiv:1602.07360},
  year={2016}
}

Usage

$CAFFE_ROOT/build/tools/caffe test --model=trainval.prototxt --weights=SqueezeNet_residual_top1_0.6038_top5_0.8250.caffemodel --iterations=1000 --gpu 0

Result

I0422 14:07:39.810755 32299 caffe.cpp:293] accuracy_top1 = 0.603759
I0422 14:07:39.810775 32299 caffe.cpp:293] accuracy_top5 = 0.824981
I0422 14:07:39.810792 32299 caffe.cpp:293] loss = 1.76711 (* 1 = 1.76711 loss) 

Architecture of the residual SqueezeNet


The building block:

Releases

No releases published

Packages

No packages published