This repository contains a MXNet implementation of the paper Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution.

Model Top1
ResNet-v1-50 76.05
OctResNet-v1-50 77.47
OctResNet-v1-50-cosine 78.04

OctResNet-v1-50-cosine model used alpha = 0.25 in the table 2 of the paper.

To Do List

  • [ ] support mobilenet v1/v2


This repo is based on DPN.


Here is the official implementation: https://github.com/facebookresearch/OctConv (Highly Recommend!)