def test_can_reconstruct_standard_encoding_with_recursive_mixing_alg(self):
kappa = [1, 1]
delta = [2, 1]
b = 1.5
int_shift = [-1, -0.1]
omega = np.pi
delta_t = 1e-4
t = np.arange(0, 25, delta_t)
np.random.seed(10)
original = bandlimitedSignal(t, delta_t, omega)
y = original.sample(t)
y = np.atleast_2d(y)
A = [[1], [2]]
tem_mult = timeEncoder(kappa, delta, b, A, integrator_init=int_shift)
spikes_mult = tem_mult.encode(y, delta_t)
rec_mult = tem_mult.decode_recursive(spikes_mult,
t,
omega,
delta_t,
num_iterations=20)
start_index = int(y.shape[1] / 10)
end_index = int(y.shape[1] * 9 / 10)
assert (np.mean(
((rec_mult[0, :] - y[0, :])**2)[start_index:end_index]) /
np.mean(y[0, :]**2) < 1e-3)
def test_can_reconstruct_precise_encoding_with_3_by_2_mixing_recursive(
self):
kappa = [1, 1, 1]
delta = [2, 1, 1]
b = [2, 2, 2]
int_shift = [-1, -0.1, 0.2]
omega = np.pi
delta_t = 1e-4
end_time = 25
t = np.arange(0, 25, delta_t)
y_param = []
np.random.seed(10)
original1 = bandlimitedSignal(t, delta_t, omega)
np.random.seed(11)
original2 = bandlimitedSignal(t, delta_t, omega)
y = np.zeros((2, len(t)))
y[0, :] = original1.sample(t)
y[1, :] = original2.sample(t)
y = np.atleast_2d(y)
A = [[0.9, 0.1], [0.2, 0.8], [1, 1]]
y_param.append(original1)
y_param.append(original2)
tem_mult = timeEncoder(kappa, delta, b, A, integrator_init=int_shift)
spikes_mult = tem_mult.encode_precise(
y_param,
omega,
end_time,
)
rec_mult = tem_mult.decode_recursive(spikes_mult,
t,
omega,
delta_t,
num_iterations=20)
start_index = int(y.shape[1] / 10)
end_index = int(y.shape[1] * 9 / 10)
assert (np.mean(
((rec_mult[0, :] - y[0, :])**2)[start_index:end_index]) /
np.mean(y[0, :]**2) < 1e-3)
assert (np.mean(
((rec_mult[1, :] - y[1, :])**2)[start_index:end_index]) /
np.mean(y[0, :]**2) < 1e-3)
def test_can_reconstruct_standard_encoding(self):
kappa = 1
delta = 1
omega = np.pi
delta_t = 1e-4
t = np.arange(0, 15, delta_t)
np.random.seed(10)
original = bandlimitedSignal(t, delta_t, omega)
y = original.sample(t)
b = np.max(np.abs(y)) + 1
tem_single = timeEncoder(kappa, delta, b, mixing_matrix=[[1]])
spikes_single = tem_single.encode(y, delta_t)
rec_single = tem_single.decode(spikes_single, t, omega, delta_t)
start_index = int(len(y) / 10)
end_index = int(len(y) * 9 / 10)
assert (np.mean(
((rec_single - y)**2)[start_index:end_index]) / np.mean(y**2) <
1e-3)
def test_can_compute_q_matrix(self):
kappa = [1, 1]
delta = [2, 1]
b = 2
int_shift = [-1, -0.1]
omega = np.pi
delta_t = 1e-4
t = np.arange(0, 25, delta_t)
np.random.seed(10)
original = bandlimitedSignal(t, delta_t, omega)
y = original.sample(t)
y = np.atleast_2d(y)
A = [[1], [2]]
# Compute q matrix in 2 ways and make sure they match
tem_mult = timeEncoder(kappa, delta, b, A, integrator_init=int_shift)
spikes_mult = tem_mult.encode(y, delta_t)
q1, G = tem_mult.get_closed_form_matrices(spikes_mult, omega)
q2 = tem_mult.get_integrals(
np.array(A).dot(y), spikes_mult, delta_t, q1.shape)
assert np.mean((q1 - q2)**2) < (delta_t)**2
def test_TEM_can_reconstruct_standard_encoding_ex2(self):
kappa = [3, 3, 3, 3]
delta = [1, 1, 1, 1]
int_shift = [-1, -0.5, 0, 0.5]
omega = np.pi
delta_t = 1e-4
t = np.arange(0, 20, delta_t)
np.random.seed(10)
original = bandlimitedSignal(t, delta_t, omega)
y = original.sample(t)
b = np.max(np.abs(y)) + 1
tem_mult = timeEncoder(kappa,
delta,
b,
mixing_matrix=[[1]] * 4,
integrator_init=int_shift)
spikes_mult = tem_mult.encode(y, delta_t)
rec_mult = tem_mult.decode(spikes_mult, t, omega, delta_t)
start_index = int(len(y) / 10)
end_index = int(len(y) * 9 / 10)
assert (np.mean(
((rec_mult - y)**2)[start_index:end_index]) / np.mean(y**2) < 1e-3)