rhythm92 / chainer-resdrop

Deep Networks with Stochastic Depth implementation by Chainer

Deep Networks with Stochastic Depth implementation by Chainer

Implementation by Chainer. Original paper is Deep Networks with Stochastic Depth.

This repository includes network definition scripts only.

If you want to train ResDrop from scratch, see chainer sample code.

Requirements

• Chainer 1.5+ (Neural network framework)

Usage

In python script, write:

from ResDrop152 import ResNet
model = ResNet()

Traning speed

About 25% faster per iteration than ResNet with no layer drop.

Sample result

I trained ResNet101 with layer drop and ResNet101 with no layer drop for PASCAL VOC Action dataset. ResNet with layer drop improved the accuracy of test results about 4%.