Learning-and-Intelligent-Systems / Analytical-Learning-Theory

97.39% on CIFAR10 with PyTorch

This repository contains a PyTorch implementation code for reproducing the results in our paper:

Generalization in Machine Learning via Analytical Learning Theory
Kenji Kawaguchi, Yoshua Bengio, Vikas Verma, and Leslie Pack Kaelbling

Test error (%) with WideResNet28_10 and different regularization methods

Regularization Method	CIFAR-10	CIFAR-100	SVHN
Standard	3.79 ± 0.07	19.85 ± 0.14	2.47 ± 0.04
Single-cutout	3.19 ± 0.09	18.13 ± 0.28	2.23 ± 0.03
Dual-cutout	2.61 ± 0.04	17.54 ± 0.09	2.06 ± 0.06

• Dual-cutout is proposed in our paper based on a new learning theory.

How to run DualCutout

python cifar10/resnext/main.py --dualcutout --dataset cifar10 --arch wrn28_10 \
--epochs 300 --batch_size 64 --learning_rate 0.1 --data_aug 1 --decay 0.0005 --schedule 150 225 \
--gamma 0.1 0.1 --alpha 0.1 --cutsize 16

Add the --temp_dir and --home_dir as appropriate in the above commands. For Cifar10 and Cifar100, we used --cutsize 16, and for SVHN, we used --cutsize 20.

How to run Single Cutout

python cifar10/resnext/main.py --singlecutout --dataset cifar10 --arch wrn28_10 \
--epochs 300 --batch_size 64 --learning_rate 0.1 --data_aug 1 --decay 0.0005 --schedule 150 225 \
--gamma 0.1 0.1 --alpha 0.1 --cutsize 16

How to run baseline

python cifar10/resnext/main.py --dataset cifar10 --arch wrn28_10 \
--epochs 300 --batch_size 64 --learning_rate 0.1 --data_aug 1 --decay 0.0005 --schedule 150 225 \
--gamma 0.1 0.1

This code has been tested with
python 2.7.9
torch 0.3.1
torchvision 0.2.0