Python get_dkl_model Examples

Example #1

    'CIFAR10 - LSUN': [],
    'CIFAR10 - FarAway': [],
    'CIFAR10 - Adversarial': [],
    'CIFAR10 - FarAwayAdv': []
}

tab_cal = {'DKL': ([], [])}

delta = 2000

np.random.seed(args.randseed)
torch.manual_seed(args.randseed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False

model_dkl, likelihood = dkl.get_dkl_model(dataset='CIFAR10')

model_dkl.cuda()
likelihood.cuda()

state = torch.load(f'./pretrained_models/CIFAR10_dkl.pt')
model_dkl.load_state_dict(state['model'])
likelihood.load_state_dict(state['likelihood'])
model_dkl.eval()
likelihood.eval()

# In-distribution
py_in, time_pred = timing(
    lambda: predict_dkl(test_loader, model_dkl, likelihood).cpu().numpy())
acc_in = np.mean(np.argmax(py_in, 1) == targets)
conf_in = get_confidence(py_in)

Example #2

    'MNIST - FMNIST': [],
    'MNIST - FarAway': [],
    'MNIST - Adversarial': [],
    'MNIST - FarAwayAdv': []
}

tab_cal = {'DKL': ([], [])}

delta = 2000

np.random.seed(args.randseed)
torch.manual_seed(args.randseed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False

model_dkl, likelihood = dkl.get_dkl_model(dataset='MNIST')

model_dkl.cuda()
likelihood.cuda()

state = torch.load(f'./pretrained_models/MNIST_dkl.pt')
model_dkl.load_state_dict(state['model'])
likelihood.load_state_dict(state['likelihood'])
model_dkl.eval()
likelihood.eval()

# In-distribution
py_in, time_pred = timing(
    lambda: predict_dkl(test_loader, model_dkl, likelihood).cpu().numpy())
acc_in = np.mean(np.argmax(py_in, 1) == targets)
conf_in = get_confidence(py_in)

Example #3

    'SVHN - LSUN': [],
    'SVHN - FarAway': [],
    'SVHN - Adversarial': [],
    'SVHN - FarAwayAdv': []
}

tab_cal = {'DKL': ([], [])}

delta = 2000

np.random.seed(args.randseed)
torch.manual_seed(args.randseed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False

model_dkl, likelihood = dkl.get_dkl_model(dataset='SVHN')

model_dkl.cuda()
likelihood.cuda()

state = torch.load(f'./pretrained_models/SVHN_dkl.pt')
model_dkl.load_state_dict(state['model'])
likelihood.load_state_dict(state['likelihood'])
model_dkl.eval()
likelihood.eval()

# In-distribution
py_in, time_pred = timing(
    lambda: predict_dkl(test_loader, model_dkl, likelihood).cpu().numpy())
acc_in = np.mean(np.argmax(py_in, 1) == targets)
conf_in = get_confidence(py_in)

Example #4

File: train_dkl.py Project: redknightlois/last_layer_laplace

elif args.dataset == 'CIFAR10':
    train_loader = dl.CIFAR10(train=True)
    val_loader, test_loader = dl.CIFAR10(train=False, augm_flag=False, val_size=2000)
elif args.dataset == 'SVHN':
    train_loader = dl.SVHN(train=True)
    val_loader, test_loader = dl.SVHN(train=False, augm_flag=False, val_size=2000)
elif args.dataset == 'CIFAR100':
    train_loader = dl.CIFAR100(train=True)
    val_loader, test_loader = dl.CIFAR100(train=False, augm_flag=False, val_size=2000)

targets = torch.cat([y for x, y in test_loader], dim=0).numpy()
targets_val = torch.cat([y for x, y in val_loader], dim=0).numpy()

print(len(train_loader.dataset), len(val_loader.dataset), len(test_loader.dataset))

model, likelihood = dkl.get_dkl_model(dataset=args.dataset)

model = model.cuda()
likelihood = likelihood.cuda()


if args.dataset == 'MNIST':
    lr = 1e-3
    opt = torch.optim.Adam([
        {'params': model.feature_extractor.parameters(), 'weight_decay': 5e-4},
        {'params': model.gp_layer.hyperparameters(), 'lr': lr},
        {'params': model.gp_layer.variational_parameters()},
        {'params': likelihood.parameters()},
    ], lr=lr, weight_decay=0)
else:
    lr = 0.1