vat_tf

Tensorflow implementation for reproducing the semi-supervised learning results on SVHN and CIFAR-10 dataset in the paper "Virtual Adversarial Training: a Regularization Method for Supervised and Semi-Supervised Learning" http://arxiv.org/abs/1704.03976

Requirements

tensorflow-gpu 1.1.0, scipy 0.19.0(for ZCA whitening)

Preparation of dataset for semi-supervised learning

On CIFAR-10

python cifar10.py --data_dir=./dataset/cifar10/

On SVHN

python svhn.py --data_dir=./dataset/svhn/

Semi-supervised Learning without augmentation

On CIFAR-10

python train_semisup.py --dataset=cifar10 --data_dir=./dataset/cifar10/ --log_dir=./log/cifar10/ --num_epochs=500 --epoch_decay_start=460 --epsilon=10.0 --method=vat

On SVHN

python train_semisup.py --dataset=svhn --data_dir=./dataset/svhn/ --log_dir=./log/svhn/ --num_epochs=120 --epoch_decay_start=80 --epsilon=2.5 --top_bn --method=vat

Semi-supervised Learning with augmentation

On CIFAR-10

python train_semisup.py --dataset=cifar10 --data_dir=./dataset/cifar10/ --log_dir=./log/cifar10aug/ --num_epochs=500 --epoch_decay_start=460 --aug_flip=True --aug_trans=True --epsilon=8.0 --method=vat

On SVHN

python train_semisup.py --dataset=svhn --data_dir=./dataset/svhn/ --log_dir=./log/svhnaug/ --num_epochs=120 --epoch_decay_start=80 --epsilon=3.5 --aug_trans=True --top_bn --method=vat

Semi-supervised Learning with augmentation + entropy minimization

On CIFAR-10

python train_semisup.py --dataset=cifar10 --data_dir=./dataset/cifar10/ --log_dir=./log/cifar10aug/ --num_epochs=500 --epoch_decay_start=460 --aug_flip=True --aug_trans=True --epsilon=8.0 --method=vatent

On SVHN

python train_semisup.py --dataset=svhn --data_dir=./dataset/svhn/ --log_dir=./log/svhnaug/ --num_epochs=120 --epoch_decay_start=80 --epsilon=3.5 --aug_trans=True --top_bn --method=vatent

Evaluation of the trained model

On CIFAR-10

python test.py --dataset=cifar10 --data_dir=./dataset/cifar10/ --log_dir=<path_to_log_dir>

On SVHN

python test.py --dataset=svhn --data_dir=./dataset/svhn/ --log_dir=<path_to_log_dir> --top_bn