PolyNet: A Pursuit of Structural Diversity in Very Deep Networks

By Xingcheng Zhang, Zhizhong Li, Chen Change Loy, Dahua Lin

Multimedia Laboratory, The Chinese University of Hong Kong

BibTeX

@article{zhang2016polynet,
  title={Polynet: A pursuit of structural diversity in very deep networks},
  author={Zhang, Xingcheng and Li, Zhizhong and Loy, Chen Change and Lin, Dahua},
  journal={arXiv preprint arXiv:1611.05725},
  year={2016}
}

The Very Deep PolyNet

• Visualization

• Models

  • Parrots: model files to be added;
  • Caffe: proto, model.

  NOTE: The model is trained using our own deep learning framework Parrots. The caffe model is converted from the Parrots model, and the proto file is based on Yuanjun's fork of Caffe.

Results

model	training speed* (#imgs/second)	single-crop val top-1	single-crop val top-5	single-crop test top-5	multi-crop val top-1	multi-crop val top-5
ResNet-152	-	22.16	6.16	-	19.38	4.49
ResNet-152^	279 ( 8 GPUs)	20.93	5.54	5.50	18.50	3.97
ResNet-269^	245 (16 GPUs)	19.78	4.89	4.82	17.54	3.55
ResNet-500^	248 (32 GPUs)	19.66	4.78	4.70	17.59	3.63
Inception-v4	-	20.0	5.0	-	1 7.7	3.8
Inception-ResNet-v2	-	19.9	4.9	-	17.8	3.7
Inception-ResNet-v2	314 ( 8 GPUs)	20.05	5.05	5.11	18.41	3.98
Very Deep Inception-ResNet	278 (32 GPUs)	19.10	4.48	4.46	17.39	3.56
Very Deep PolyNet	290 (32 GPUs)	18.71	4.25	4.33	17.36	3.45

^ The ResNet models are trained by Tong Xiao;

* Training speed is measured on Parrots using NVIDIA TITAN X Graphics Cards.