xushu-me / octconv.pytorch

PyTorch implementation of Octave Convolution with pre-trained Oct-ResNet and Oct-MobileNet models

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

Architecture	LR decay strategy	Parameters	GFLOPs	Top-1 / Top-5 Accuracy (%)
ResNet-50	step (90 epochs)	25.557M	4.089	76.010 / 92.834
ResNet-50	cosine (120 epochs)	25.557M	4.089	77.150 / 93.468
Oct-ResNet-50 (alpha=0.5)	cosine (120 epochs)	25.557M	2.367	77.640 / 93.662
ResNet-101	cosine (120 epochs)	44.549M	7.801	78.898 / 94.304
Oct-ResNet-101 (alpha=0.5)	cosine (120 epochs)	44.549M	3.991	78.794 / 94.330
ResNet-152	cosine (120 epochs)	60.193M	11.514	79.234 / 94.556
Oct-ResNet-152 (alpha=0.5)	cosine (120 epochs)	60.193M	5.615	79.258 / 94.480

MobileNet V1 on ImageNet

Architecture	LR decay strategy	Parameters	FLOPs	Top-1 / Top-5 Accuracy (%)
MobileNetV1	cosine (150 epochs)	4.232M	568.7M	72.238 / 90.536
Oct-MobileNetV1	cosine (150 epochs)	4.232M	318.2M	71.254 / 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}
}