def get_conf(dataloader, delta, model, T=1, laplace=True):
if laplace:
py = llla.predict(dataloader,
model,
M_W,
M_b,
U,
V,
B,
delta=delta)
else:
py = predict(dataloader, model, delta=delta, T=T)
conf = get_confidence(py.cpu().numpy())
return conf
var0 = torch.tensor(0.2656).float().cuda()
print(var0)
M_W, M_b, U, V, B = llla.estimate_variance(var0, hessians)
np.save(f'./pretrained_models/CIFAR10_{args.type}_llla.npy',
[M_W, M_b, U, V, B])
else:
time_inf = 0
M_W, M_b, U, V, B = list(
np.load(f'./pretrained_models/CIFAR10_{args.type}_llla.npy',
allow_pickle=True))
# In-distribution
py_in, time_pred = timing(
lambda: llla.predict(test_loader, model, M_W, M_b, U, V, B).cpu().numpy())
acc_in = np.mean(np.argmax(py_in, 1) == targets)
conf_in = get_confidence(py_in)
mmc = conf_in.mean()
ece, mce = get_calib(py_in, targets)
save_res_ood(tab_ood['CIFAR10 - CIFAR10'], mmc)
save_res_cal(tab_cal['LLLA'], ece, mce)
print(
f'[In, LLLA] Time: {time_inf:.1f}/{time_pred:.1f}s; Accuracy: {acc_in:.3f}; ECE: {ece:.3f}; MCE: {mce:.3f}; MMC: {mmc:.3f}'
)
# Out-distribution
py_out = llla.predict(test_loader_SVHN, model, M_W, M_b, U, V, B).cpu().numpy()
conf_svhn = get_confidence(py_out)
mmc = conf_svhn.mean()
auroc = get_auroc(py_in, py_out)