print("test MSE of zero function", np.mean(ytest ** 2))
print('\n\n','*'*80,'\nrun',run)
for num_examples in L_num_examples:
print('______\nsample size:', num_examples)
print('Current Experiment: Addition with noise '+str(noise_level) +' and ' + str(num_states)+' states')
data_function = lambda l: generate_data_simple_addition(num_examples, l, noise_level =noise_level)
Xl, yl = data_function(length)
X2l, y2l = data_function(length * 2)
X2l1, y2l1 = data_function(length * 2 + 1)
for method in methods:
print(method)
if method != 'LSTM' and method != 'TIHT+SGD' and method != 'ALS':
Tl = learning.sequence_to_tensor(Xl)
T2l = learning.sequence_to_tensor(X2l)
T2l1 = learning.sequence_to_tensor(X2l1)
t=tic()
Hl = learning.approximate_hankel(Tl, yl, alpha_ini_value=alpha,
rank=num_states, eps=TIHT_epsilon,
learning_rate=TIHT_learning_rate, max_iters=TIHT_max_iters,
method=method, verbose=-1)
H2l = learning.approximate_hankel(T2l, y2l, alpha_ini_value=alpha,
rank=num_states, eps=TIHT_epsilon,
learning_rate=TIHT_learning_rate, max_iters=TIHT_max_iters,
method=method, verbose=-1)
H2l1 = learning.approximate_hankel(T2l1, y2l1, alpha_ini_value=alpha, rank=num_states, eps=TIHT_epsilon,
learning_rate=TIHT_learning_rate, max_iters=TIHT_max_iters,
method=method, verbose=-1)
X_temp = np.swapaxes(X_temp, 1, 2)
# print(X_temp[0:5])
# print(Y_temp[0:5])
# Tl = learning.sequence_to_tensor(X_temp)
#
# H_temp = learning.approximate_hankel(Tl, Y_temp, alpha_ini_value=1.,
# rank=num_states, eps=TIHT_epsilon,
# learning_rate=TIHT_learning_rate,
# max_iters=TIHT_max_iters,
# method='TIHT', verbose=-1)
# print(H_temp, H_temp.shape)
X_temp, Y_temp = synthetic_data.pad_data(
X_temp_vec, Y_temp_vec)
X_temp = np.swapaxes(X_temp, 1, 2)
Tl = learning.sequence_to_tensor(X_temp)
H_temp = learning.approximate_hankel(
Tl,
Y_temp,
alpha_ini_value=1.,
rank=num_states,
eps=TIHT_epsilon,
learning_rate=TIHT_learning_rate,
max_iters=TIHT_max_iters,
method='TIHT',
verbose=-1)
#print(H_temp)
H_temp = H_temp.squeeze()
#H_temp = recover_tensor(H_temp).squeeze()
H_temp_shape = list(H_temp.shape)
alpha=1.,
lifting=True)
Xtest_temp = torch.from_numpy(Xtest).float()
pred = model(Xtest_temp)
pred_numpy = pred.detach().numpy().reshape(-1, )
pred_numpy = pred_numpy.reshape(ytest.shape)
rmse1 = cal_RMSE(pred_numpy, ytest, mean_data, std_data)
mape1 = cal_MAPE(pred_numpy, ytest, mean_data, std_data)
mae1 = cal_MAE(pred_numpy, ytest, mean_data, std_data)
else:
X_vec, y_vec = [], []
for i in range(0, len(X)):
X_vec.append(learning.sequence_to_tensor(X[i]))
y_vec.append(Y[i])
TL_vec = learning.get_all_TLs(X_vec,
y_vec,
rank=num_states,
eps=TIHT_epsilon,
learning_rate=TIHT_learning_rate,
max_iters=TIHT_max_iters,
method=method,
verbose=0,
alpha_ini_value=1.)
Hl = learning.approximate_hankel_l(TL_vec=TL_vec, length=length)
Hl_plus = learning.approximate_hankel_plus(TL_vec=TL_vec,
length=length)
Hl_minus = learning.approximate_hankel_minus(TL_vec=TL_vec,