AgnosticMC

Environment:

• Python 3.5.3
• PyTorch 0.4.1

Dataset

CIFAR10 and MNIST datasets are required here. When you first run the script below, if you haven't downloaded the data before, it will automatically download them at folder data_CIFAR10 and data_MNIST. Stay patient.

Script:

First, please go to the Bin_CIFAR10 folder. For now, we only update the codes in that case. There are a lot of loss terms, described in the argparse part, if the loss weight (shorted as lw in the code) is 0, it means this loss term will not be included in the total loss.

test DFL using random noise as input on MNIST:

# debug mode, print log on screen
python main.py  --lw_soft 100  --lw_hard_dec 0  --lw_hard_se 0  --use_random_input  --gpu <id>  

# formal experiment mode, print log in backend
nohup python main.py  --lw_soft 100  --lw_hard_dec 0  --lw_hard_se 0  --use_random_input  --gpu <id> --CodeID <code git log id>  > /dev/null &

# newest script:
nohup python main.py  --lw_soft 100  --lw_hard_se 0  --num_epoch 400 --gpu <id>  --lw_class_balance 0  --lw_msgan 10 --lw_msgan_feat 0 --use_condition --CodeID <code git log id> > /dev/null &

The log will be saved in folder ../Experiment/xxx, where xxx is a folder named by the time stamp when you run the code. For details, please refer to the set_up_dir function in util.py. An example for xxx: SERVER218-20190501-124737_test. It tells you where and when you run the experiment, easy for later check.