A MXNet Implementation for Drop an Octave
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.
- support mobilenet v1/v2
This repo is based on DPN.
Here is the official implementation: https://github.com/facebookresearch/OctConv (Highly Recommend!)
