Home

Awesome

octconv.pytorch

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

ResNet-50/101 on ImageNet

ArchitectureLR decay strategyParametersGFLOPsTop-1 / Top-5 Accuracy (%)
ResNet-50step (90 epochs)25.557M4.08976.010 / 92.834
ResNet-50cosine (120 epochs)25.557M4.08977.150 / 93.468
Oct-ResNet-50 (alpha=0.5)cosine (120 epochs)25.557M2.36777.640 / 93.662
ResNet-101cosine (120 epochs)44.549M7.80178.898 / 94.304
Oct-ResNet-101 (alpha=0.5)cosine (120 epochs)44.549M3.99178.794 / 94.330
ResNet-152cosine (120 epochs)60.193M11.51479.234 / 94.556
Oct-ResNet-152 (alpha=0.5)cosine (120 epochs)60.193M5.61579.258 / 94.480
<p align="center"><img src="fig/ablation.png" width="600" /></p>

MobileNet V1 on ImageNet

ArchitectureLR decay strategyParametersFLOPsTop-1 / Top-5 Accuracy (%)
MobileNetV1cosine (150 epochs)4.232M568.7M72.238 / 90.536
Oct-MobileNetV1cosine (150 epochs)4.232M318.2M71.254 / 89.728

Acknowledgement

Official MXNet implmentation by @cypw

Citation

@InProceedings{Chen_2019_ICCV,
author = {Chen, Yunpeng and Fan, Haoqi and Xu, Bing and Yan, Zhicheng and Kalantidis, Yannis and Rohrbach, Marcus and Yan, Shuicheng and Feng, Jiashi},
title = {Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks With Octave Convolution},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}