SP_CUP

1. Load Data

If you are using the cup training dataset and validation dataset, please ensure to load the following files in your code:

• cup_train.h5: Contains the training dataset
• cup_val.h5: Contains the validation dataset

parser.add_argument('--train_datapath', default='cup_train.csv', type=str)
parser.add_argument("--test_datapath", type=str,
                        default='cup_val.csv', help="test data path")

    train_dataset = ImageDataset(
    hdf5_filename='cup_train.h5',
    labels_filename=args.train_datapath,
    dataset_name='train_features'
)


    test_dataset = ImageDataset(
    hdf5_filename='cup_val.h5',
    labels_filename=args.test_datapath,
    dataset_name='test_features'
)

If you are using the AI-Face training dataset and CUP validation dataset, please ensure to load the following files in your code:

• ai_face_all_v1.h5: Contains the AI-Face training dataset
• cup_val.h5: Contains the validation dataset

parser.add_argument('--train_datapath', default='all_ai_face_v1.csv', type=str)
parser.add_argument("--test_datapath", type=str,
                        default='cup_val.csv', help="test data path")

    train_dataset = ImageDataset(
    hdf5_filename='cup_train.h5',
    labels_filename=args.train_datapath,
    dataset_name='train_features'
)


    test_dataset = ImageDataset(
    hdf5_filename='cup_val.h5',
    labels_filename=args.test_datapath,
    dataset_name='test_features'
)

2. Train the model

CVaR Loss

• If you are using CVaR loss, change the loss type to cvar, and tune the alpha from [0.1, 0.9]

python model_trainer(loss_type='cvar',alpha=0.7,gamma=0.8)

AUC Loss

• If you are using AUC loss, change the loss type to auc, and tune the gamma from [0.1, 0.9]

python model_trainer(loss_type='auc',alpha=0.7,gamma=0.8)

VS Loss

• If you are using VS loss, change the loss type to vs.

python model_trainer(loss_type='vs',alpha=0.7,gamma=0.8)

In VS loss you should tune hyperparameters:iota_pos,iota_neg,Delta_pos,Delta_neg. You can consider the iota [-1,1], delta [0.5,2].

elif loss_type == 'vs':
    vs_loss_fn = BinaryVSLoss(iota_pos=-0.05, iota_neg=0.05, Delta_pos=0.9, Delta_neg=1.1, weight=None)
    loss = vs_loss_fn(output,labels)
    return loss