Compressed networks from ENC (Caffe model)

Fine-tuned network models by "Efficient Neural Network Compression", CVPR 2019.

For the source-code of paper, please refer to [ENC]

• This repository contains the prototxt files
• Trained models: [driver]

AlexNet with ImageNet

• Table 1

Method	FLOPs	Weights	Top-1 Acc.	Top-5 Acc.
[ENC-Inf]	37.5%	18.0%	56.74%	80.14%
[ENC-Model]	37.5%	18.0%	56.71%	80.13%

• Table 3

Method	FLOPs	Top-1 Acc.	Top-5 Acc.
[ENC-Inf]	31%	56.66%	79.74%

• Fig.7(a) - prototxt

Method	FLOPs	Top-1 Acc.	Top-5 Acc.
[ENC-Inf]	31%	56.66%	79.74%
[ENC-Inf]	50%	57.33%	80.33%
[ENC-Inf]	75%	57.67%	80.50%
[ENC-Inf]	95%	57.74%	80.57%

• Fig.7(b) - prototxt

Method	FLOPs	Weights	Top-1 Acc.	Top-5 Acc.
[ENC-Model]	23%	18%	54.48%	78.58%
[ENC-Model]	25%	18%	55.08%	78.99%
[ENC-Model]	30%	18%	56.12%	79.59%

VGG-16 with ImageNet

• Table 1

Method	FLOPs	Top-1 Acc.	Top-5 Acc.
[ENC-Inf]	25%	71.29%	90.12%
[ENC-Model]	25%	71.25%	90.12%
[ENC-Map]	25%	70.90%	89.97%

• Table 2

Method	FLOPs	Top-1 Acc.	Top-5 Acc.
[ENC-Model]	20%	71.06%	89.95%

• Table 3

Method	FLOPs	Top-1 Acc.	Top-5 Acc.
[ENC-Model]	24%	70.95%	89.95%

ResNet-56 with Cifar10

• Table 1

Method	FLOPs	Top-1 Acc. w/o FT	Top-1 Acc. w/ FT
[ENC-Inf]	50%	90.22%	93.0%
[ENC-Model]	50%	89.55%	93.0%
[ENC-Map]	50%	89.80%	93.0%

• Table 3

Method	FLOPs	Top-1 Acc.
[ENC-Map]	55%	93.2%

Citation

@CONFERENCE{ENC_CVPR19,
  author={Hyeji Kim, Muhammad Umar Karim Khan, Chong-Min Kyung},
  title={Efficient Neural Network Compression},
  booktitle={CVPR},
  month = {June},
  year = {2019},
}