def nested_stf(dataset, cur_fold, r, lr, num_iter):
train, test = get_train_test(dataset, num_folds=num_folds, fold_num=cur_fold)
#train, valid = train_test_split(train, test_size=0.2, random_state=0)
valid = train[int(0.8*len(train)):].copy()
train = train[:int(0.8 * len(train))].copy()
valid_gt = valid[:, 1:, :, :]
test_gt = test[:, 1:, :, :]
valid_copy = valid.copy()
valid_copy[:, 1:, :, :] =np.NaN
train_valid = np.concatenate([train, valid_copy])
H, A, D, T = stf_4dim_time(tensor=train_valid, r=r, lr=lr, num_iter=num_iter)
valid_pred = np.einsum("Hr, Ar, Dr, ATr -> HADT", H, A, D, T)[len(train):, 1:, :, :]
valid_error = {APPLIANCE_ORDER[i+1]:mean_absolute_error(valid_pred[:, i,:,:].flatten(),
valid_gt[:, i, :, :].flatten()) for i in range(valid_pred.shape[1])}
test_copy = test.copy()
test_copy[:, 1:, :, :] =np.NaN
train_test = np.concatenate([train, test_copy])
H, A, D, T = stf_4dim_time(tensor=train_test, r=r, lr=lr, num_iter=num_iter)
test_pred = np.einsum("Hr, Ar, Dr, ATr -> HADT", H, A, D, T)[len(train):, 1:, :, :]
test_error = {APPLIANCE_ORDER[i+1]:mean_absolute_error(test_pred[:, i,:,:].flatten(),
test_gt[:, i, :, :].flatten()) for i in range(test_pred.shape[1])}
return valid_pred, valid_error, valid_gt, test_pred, test_error, test_gt
def mtf(r=2, lr=1, num_iter=100):
out = []
for cur_fold in range(5):
train, test = get_train_test(num_folds=num_folds, fold_num=cur_fold)
test_copy = test.copy()
test_copy[:, 1:, :, :] = np.NaN
train_test = np.concatenate([train, test_copy])
H, A, D, T = stf_4dim_time(tensor=train_test,
r=r,
lr=lr,
num_iter=num_iter)
pred = np.einsum("Hr, Ar, Dr, ATr ->HADT", H, A, D, T)[len(train):,
1:, :, :]
out.append(pred)
return np.concatenate(out)