vat_tf

by takerum

takerum / vat_tf

Virtual adversarial training with Tensorflow

217 Stars 73 Forks Last release: Not found MIT License 36 Commits 0 Releases

Available items

No Items, yet!

The developer of this repository has not created any items for sale yet. Need a bug fixed? Help with integration? A different license? Create a request here:

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=

On SVHN

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

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.