Skip to content

MTL-IBP

MTL-IBP compatibility entry point for the common expressive IBP trainer.

mtl_ibp_train_model(original_model, hardened_model, train_loader, val_loader=None, start_epoch=0, end_epoch=None, num_epochs=None, eps=0.3, eps_std=0.3, eps_schedule=(0, 20, 50), eps_schedule_unit='epoch', eps_scheduler_args=dict(), optimizer=None, lr_scheduler=None, lr_decay_schedule_unit='epoch', n_classes=10, gradient_clip=None, shi_regularisation_weight=0.5, shi_reg_decay=True, l1_regularisation_weight=1e-05, alpha=0.5, pgd_restarts=1, pgd_step_size=10, pgd_n_steps=1, pgd_eps_factor=1, pgd_decay_factor=0.1, pgd_decay_checkpoints=(), pgd_early_stopping=False, results_path='./results', checkpoint_save_interval=10, device='cuda', relu_upper_retention=1.0, population_bn=False, loss_type='mtl_ibp', validation_eps=None, selection_state=None, *, history=None)

Train MTL-IBP using the shared expressive IBP training procedure.

Arguments follow expressive_ibp_train_model; this entry point selects the method-specific loss while reusing its schedules and training loop.

Source code in CTRAIN/train/certified/mtl_ibp.py 
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
def mtl_ibp_train_model(
    original_model,
    hardened_model,
    train_loader,
    val_loader=None,
    start_epoch=0,
    end_epoch=None,
    num_epochs=None,
    eps=0.3,
    eps_std=0.3,
    eps_schedule=(0, 20, 50),
    eps_schedule_unit="epoch",
    eps_scheduler_args=dict(),
    optimizer=None,
    lr_scheduler=None,
    lr_decay_schedule_unit="epoch",
    n_classes=10,
    gradient_clip=None,
    shi_regularisation_weight=0.5,
    shi_reg_decay=True,
    l1_regularisation_weight=0.00001,
    alpha=0.5,
    pgd_restarts=1,
    pgd_step_size=10,
    pgd_n_steps=1,
    pgd_eps_factor=1,
    pgd_decay_factor=0.1,
    pgd_decay_checkpoints=(),
    pgd_early_stopping=False,
    results_path="./results",
    checkpoint_save_interval=10,
    device="cuda",
    relu_upper_retention=1.0,
    population_bn=False,
    loss_type='mtl_ibp', validation_eps=None, selection_state=None, *, history=None):
    """Train MTL-IBP using the shared expressive IBP training procedure.

    Arguments follow expressive_ibp_train_model; this entry point selects the
    method-specific loss while reusing its schedules and training loop.
    """
    validate_expressive_options(alpha, loss_type)
    options = locals().copy()
    options.pop('loss_type')
    loss_func = {'mtl_ibp': get_mtl_ibp_loss, 'cc_ibp': get_cc_ibp_loss,
                 'exp_ibp': get_exp_ibp_loss}[loss_type]
    return expressive_ibp_train_model(**options, loss_func=loss_func, method=loss_type)