Learning Not to Learn in the Presence of Noisy Labels

This code runs most of the experiments detailed in the paper.

The following commands run experiments with the same hyperparameters that were used in the paper (varying the noise rate):

MNIST:

nll:

• python3 main.py --dataset mnist --noise_rate 0.2 --lr 0.001 --batch_size 128 --noise_type symmetric --n_epoch 50 --lambda_type nll

Gamblers + Early Stopping:

• python3 main.py --dataset mnist --noise_rate 0.2 --lr 0.001 --batch_size 128 --noise_type symmetric --n_epoch 50 --lambda_type gmblers --eps 9.9 --early_stopping

Gamblers + Autoscheduling:

• python3 main.py --dataset mnist --noise_rate 0.2 --lr 0.001 --batch_size 128 --noise_type symmetric --n_epoch 50 --lambda_type euc

CIFAR10:

nll:

• python3 main.py --dataset mnist --noise_rate 0.2 --lr 0.001 --batch_size 128 --noise_type symmetric --n_epoch 50 --lambda_type nll

Gamblers + Early Stopping:

• python3 main.py --dataset cifar10 --noise_rate 0.2 --lr 0.001 --batch_size 128 --noise_type symmetric --n_epoch 100 --start_gamblers 10 --lambda_type gmblers --eps 9.9 --early_stopping

Gamblers + Autoscheduling:

• python3 main.py --dataset cifar10 --noise_rate 0.2 --lr 0.001 --batch_size 128 --noise_type symmetric --n_epoch 100 --start_gamblers 10 --lambda_type euc

IMDB:

nll:

• python3 main.py --dataset imdb --noise_rate 0.2 --lr 0.001 --batch_size 32 --noise_type symmetric --n_epoch 100 --start_gamblers 10 --lambda_type nll

Gamblers + Early Stopping:

• python3 main.py --dataset imdb --noise_rate 0.2 --lr 0.001 --batch_size 32 --noise_type symmetric --n_epoch 100 --start_gamblers 10 --lambda_type gmblers --eps 1.95 --early_stopping

Gamblers + Autoscheduling:

• python3 main.py --dataset imdb --noise_rate 0.2 --lr 0.001 --batch_size 32 --noise_type symmetric --n_epoch 100 --start_gamblers 10 --lambda_type euc

By manipulating the options, the full range of experiments involving Gambler's Loss in the paper are runnable.