LRDKT

The cnn compression algorithm LRDKT.

Citation

If you find our project useful in your research, please consider citing:

@article{lin2018holistic,
  title={Holistic CNN Compression via Low-rank Decomposition with Knowledge Transfer},
  author={Lin, Shaohui and Ji, Rongrong and Chen, Chao and Tao, Dacheng and Luo, Jiebo},
  journal={IEEE transactions on pattern analysis and machine intelligence},
  year={2018},
  publisher={IEEE}
}

Caffe

We modfiy Caffe code, see the caffe directory for code and more information.

Train

See the train directory for code and more information.

Models

Test prototxt and caffemodel in test.

AlexNet

227x227 center crop validation accuracy on ImageNet, tested on one GTX TITAN X GPU with batch_size=32.

Model	#Param.	#FLOPs	CPU speedup	GPU speedup	Top-1 Err.	Top-5 Err.
LRDKT-0.7	18.7M	0.27B	2.04×	1.8×	40.89%	18.22%
LRDKT-GAP	1.1M	0.26B	2.22×	2.0×	45.28%	21.88%

VGG16

224x224 center crop validation accuracy on ImageNet, tested on one GTX TITAN X GPU with batch_size=32.

Model	#Param.	#FLOPs	CPU speedup	GPU speedup	Top-1 Err.	Top-5 Err.
LRDKT-0.7	30.5M	2.43B	2.43×	2.27×	31.16%	10.84%
LRDKT-0.5	9.5M	1.31B	2.61×	2.55×	35.77%	13.9%
LRDKT-GAP	3.3M	2.41B	2.46×	2.33×	31.84%	11.43%