def test_nbdt_gradient_soft(resnet18_cifar10, input_cifar10, label_cifar10, criterion):
output_cifar10 = resnet18_cifar10(input_cifar10)
assert output_cifar10.requires_grad
criterion = SoftTreeSupLoss(
dataset="CIFAR10", criterion=criterion, hierarchy="induced"
)
loss = criterion(output_cifar10, label_cifar10)
loss.backward()
def test_criterion_cifar10(criterion, label_cifar10):
criterion = SoftTreeSupLoss(dataset='CIFAR10',
criterion=criterion,
hierarchy='induced')
criterion(torch.randn((1, 10)), label_cifar10)
def test_criterion_tinyimagenet200(criterion):
criterion = SoftTreeSupLoss(dataset='TinyImagenet200',
criterion=criterion,
hierarchy='induced')
criterion(torch.randn((1, 200)), torch.randint(200, (1, )))
def test_criterion_cifar100(criterion):
criterion = SoftTreeSupLoss(dataset='CIFAR100',
criterion=criterion,
hierarchy='induced')
criterion(torch.randn((1, 100)), torch.randint(100, (1, )))