This repository contains a PyTorch implementation of the Multi-Sample Dropout from the paper

Multi-Sample Dropout for Accelerated Trainingand Better Generalization

by Hiroshi Inoue.

Multi-Sample Dropout is a new way to expand the traditional Dropout by using multiple dropout masks for the same mini-batch.

• PyTorch
• torchvision

The code in this repository implements Multi-Sample Dropout training, with example on the CIFAR-10 datasets.

To use Multi-Sample Dropout use the following command.

python run.py --dropout_num = 8

To experiment th result,we use CIFAR-10 dataset for MiniResNet.

# no dropout
python run.py --dropout_num=0

# sample = 1
python run.py --dropout_num=1

# sample = 8
python run.py --dropout_num=8

Train loss of Multi-Sample Dropout with MiniResNet on CIFAR-10.

Valid loss of Multi-Sample Dropout with MiniResNet on CIFAR-10.

Valid accuracy of Multi-Sample Dropout with MiniResNet on CIFAR-10.