Train network (AllCNN / LeNet) with noisy datasets, using different regularization methods. Code creates the requiered dataset, trains the network for "repeat_num" times, and saves run logs and results in a dedicated folder.
Data
| parameter | description | type | default |
|---|---|---|---|
| dataset | cifar10, cifar100 or mnist | str | 'cifar10' |
| noise_type | uniform_noise or flip_noise | str | 'uniform_noise' |
| noise_rate | rate of corrupted samples | float | 0.0 |
| validation_ratio | rate of validation data (out of train data) | float | 0.1 |
| preprocess_mode | range_0_to_1 or mean_substract | str | 'range_0_to_1' |
Train
| parameter | description | type | default |
|---|---|---|---|
| do_sn | whether or not to spectrally normalize the network weights | store_true | |
| wd_coef | coefficient of L2 loss term | float | 0.0 |
| jacob_coef | coefficient of jacobian loss term | float | 0.0 |
| entropy_coef | coefficient of entropy loss term | float | 0.0 |
| epochs_num | number of epochs in each train run | int | 30 |
| total_repeats | number of train runs | int | 5 |
| GPU_num | number of GPU to use | int | 0 |
Examples
python CodeManager.py --noise_rate 0.4 --noise_type flip_noise --dataset mnist --wd_coef 1e-4 --do_sn --GPU_num 0
python CodeManager.py --noise_rate 0.3 --noise_type uniform_noise --dataset cifar10 --wd_coef 1e-4 --do_sn --entropy_coef 1 --epochs_num 20 --GPU_num 1
python CodeManager.py --noise_rate 0.5 --noise_type flip_noise --dataset cifar100 --wd_coef 1e-4 --epochs_num 35 --GPU_num 2
python CodeManager.py --noise_rate 0.7 --noise_type uniform_noise --dataset cifar10 --wd_coef 1e-5 --jacob_coef 1e-3 --GPU_num 3