Octave Convolution (OctConv)

Pytorch implementation of Octave Convolution with other similar operation

This is third-party/un-official implementation of the following papers which are presented in Recent_Convolution.pdf:

1. Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution paper.
2. Adaptively Connected Neural Networks.(CVPR 2019) paper.
3. Res2net:A New Multi-scale Backbone Architecture paper.

Plan

• Add Res2Net block with SE-layer
• Add Adaptive-Convolution: both pixel-aware and dataset-aware (done)
• Add HetConv (optional): (if I have time 🙂)
• Train on CIFAR
• Train on ImageNet (Who can help me train this repo on ImageNet?)

Requirements

• Python 3
  • Tested with Python 3.6
• PyTorch
  • Tested with version 1.0.1

Usage

Check model files under the nn directory.

from nn.OCtaveResnet import resnet50
from nn.res2net import se_resnet50
from nn.AdaptiveConvResnet import PixelAwareResnet50, DataSetAwareResnet50

model = resnet50().cuda()
model = se_resnet50().cuda()
model = PixelAwareResnet50().cuda()
model = DataSetAwareResnet50().cuda()

Credits:

Referenced these implementations:

1. OctaveConv: MXNet implementation here
2. AdaptiveCov: Offical tensorflow implementation here

Other Implementations

• PyTorch
  • https://github.com/gan3sh500/octaveconv-pytorch
  • https://github.com/vivym/OctaveConv.pytorch
• TensorFlow / Keras
  • https://github.com/CyberZHG/keras-octave-conv
  • https://github.com/l5shi/Octave-Conv-Keras

License

MIT License