/OctConv

Code for paper

Primary LanguagePythonMIT LicenseMIT

Octave Convolution

MXNet implementation for:

Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution

ImageNet

Ablation

  • Loss: Softmax
  • Learning rate: Cosine (warm-up: 5 epochs, lr: 0.4)
  • MXNet API: Symbol API

example

Model baseline alpha = 0.125 alpha = 0.25 alpha = 0.5 alpha = 0.75
DenseNet-121 75.4 / 92.7 76.1 / 93.0 75.9 / 93.1 -- --
ResNet-26 73.2 / 91.3 75.8 / 92.6 76.1 / 92.6 75.5 / 92.5 74.6 / 92.1
ResNet-50 77.0 / 93.4 78.2 / 93.9 78.0 / 93.8 77.4 / 93.6 76.7 / 93.0
SE-ResNet-50 77.6 / 93.6 78.7 / 94.1 78.4 / 94.0 77.9 / 93.8 77.4 / 93.5
ResNeXt-50 78.4 / 94.0 -- 78.8 / 94.2 78.4 / 94.0 77.5 / 93.6
ResNet-101 78.5 / 94.1 79.2 / 94.4 79.2 / 94.4 78.7 / 94.1 --
ResNeXt-101 79.4 / 94.6 -- 79.6 / 94.5 78.9 / 94.4 --
ResNet-200 79.6 / 94.7 80.0 / 94.9 79.8 / 94.8 79.5 / 94.7 --

Note:

  • Top-1 / Top-5, single center crop accuracy is shown in the table. (testing script)
  • All residual networks in ablation study adopt pre-actice version[1] for convenience.

Others

  • Learning rate: Cosine (warm-up: 5 epochs, lr: 0.4)
  • MXNet API: Gluon API
Model alpha label smoothing[2] mixup[3] #Params #FLOPs Top1 / Top5
0.75 MobileNet (v1) .375 2.6 M 213 M 70.5 / 89.5
1.0 MobileNet (v1) .5 4.2 M 321 M 72.5 / 90.6
1.0 MobileNet (v2) .375 Yes 3.5 M 256 M 72.0 / 90.7
1.125 MobileNet (v2) .5 Yes 4.2 M 295 M 73.0 / 91.2
Oct-ResNet-152 .125 Yes Yes 60.2 M 10.9 G 81.4 / 95.4
Oct-ResNet-152 + SE .125 Yes Yes 66.8 M 10.9 G 81.6 / 95.7

Citation

@article{chen2019drop,
  title={Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution},
  author={Chen, Yunpeng and Fan, Haoqi and Xu, Bing and Yan, Zhicheng and Kalantidis, Yannis and Rohrbach, Marcus and Yan, Shuicheng and Feng, Jiashi},
  journal={Proceedings of the IEEE International Conference on Computer Vision},
  year={2019}
}

Third-party Implementations

Acknowledgement

  • Thanks MXNet, Gluon-CV and TVM!
  • Thanks @Ldpe2G for sharing the code for calculating the #FLOPs (link)
  • Thanks Min Lin (Mila), Xin Zhao (Qihoo Inc.), Tao Wang (NUS) for helpful discussions on the code development.

Reference

[1] He K, et al "Identity Mappings in Deep Residual Networks".

[2] Christian S, et al "Rethinking the Inception Architecture for Computer Vision"

[3] Zhang H, et al. "mixup: Beyond empirical risk minimization.".

License

The code and the models are MIT licensed, as found in the LICENSE file.