def sparse_algorithm(mp: np.ndarray,
Pp: np.ndarray,
model: np.ndarray,
vis: np.ndarray,
weight: np.ndarray,
Q: np.ndarray,
R: np.ndarray,
ant1: np.ndarray,
ant2: np.ndarray,
tbin_indices: np.ndarray,
tbin_counts: np.ndarray,
Nsamples: np.int32 = 10):
"""Sparse-matrix implementation of EnKF algorithm. Not
numba-compiled."""
# Time counts
n_time = len(tbin_indices)
# Number of Baselines
n_bl = model.shape[0] // n_time
# Number of Antennas
n_ant = int((np.sqrt(8 * n_bl + 1) + 1)) // 2
# Dimensions
n_chan, n_dir = model.shape[1:]
# Matrix-size
axis_length = n_ant * n_chan * n_dir
# Original matrix size
gains_shape = (n_time, n_ant, n_chan, n_dir, 2)
# Jacobian shape
shape = gains_shape[1:]
# Covariance shape
covs_shape = (n_time, 2 * axis_length, 2 * axis_length)
# State vectors
m = np.zeros(gains_shape, dtype=np.complex128)
# Covariance matrices
P = np.zeros(covs_shape, dtype=np.complex128)
# Initial state and covariance
m[0] = gains_reshape(mp, shape)
C = Pp.real
# Select CSR jacobian function
aug_jac = compute_aug_csr
# Calculate R^{-1} for a diagonal
Rinv = np.diag(1.0 / np.diag(R))
# Run Extended Kalman Filter with
# Sparse matrices
print("==> Ensemble Kalman Filter (SPARSE):")
for k in range(1, n_time):
# Progress Bar
bar_len = 100
total = n_time - 1
filled_len = int(round(bar_len * k / float(n_time - 1)))
bar = u"\u2588" * filled_len + ' '\
* (bar_len - filled_len)
print("\r%d%%|%s| %d/%d" % (k / total * 100, bar, k, total), end="")
q = process_noise(Q)
mp = gains_vector(m[k - 1])
# Ensemble Step
M = state_ensemble(mp, C) + q
mi = np.mean(M, axis=1)
# Slice indices
start = tbin_indices[k - 1]
end = start + tbin_counts[k - 1]
# Calculate Slices
row_slice = slice(start, end)
vis_slice = vis[row_slice]
model_slice = model[row_slice]
weight_slice = weight[row_slice]
ant1_slice = ant1[row_slice]
ant2_slice = ant2[row_slice]
jones_slice = gains_reshape(mi, shape)
# Calculate Augmented Jacobian
J = aug_jac(model_slice, weight_slice, jones_slice, ant1_slice,
ant2_slice)
# Hermitian of Jacobian
J_herm = J.conjugate().T
D = J @ C @ J_herm + R
K = C @ J_herm @ np.linalg.inv(D)
# Calculate Measure Vector
y = measure_vector(vis_slice, weight_slice, n_ant, n_chan)
e = measure_noise(R)
Y = np.zeros((y.shape[0], Nsamples), dtype=np.complex128)
Mf = np.zeros_like(M)
for i in range(Nsamples):
Y[:, i] = y + e[:, i]
jones_slice = gains_reshape(M[:, i], shape)
Mf = M + K @ (Y - J @ M)
mf = np.mean(Mf, axis=1)
C = np.diag(np.diag(np.cov(Mf))).real
m[k] = gains_reshape(mf, shape)
# Newline
print()
# Return Posterior states and covariances
return m
def sparse_algorithm(mp: np.ndarray, Pp: np.ndarray, model: np.ndarray,
vis: np.ndarray, weight: np.ndarray, Q: np.ndarray,
R: np.ndarray, ant1: np.ndarray, ant2: np.ndarray,
tbin_indices: np.ndarray, tbin_counts: np.ndarray):
"""Sparse-matrix implementation of EKF algorithm. Not
numba-compiled."""
# Time counts
n_time = len(tbin_indices)
# Number of Baselines
n_bl = model.shape[0] // n_time
# Number of Antennas
n_ant = int((np.sqrt(8 * n_bl + 1) + 1)) // 2
# Dimensions
n_chan, n_dir = model.shape[1:]
# Matrix-size
axis_length = n_ant * n_chan * n_dir
# Original matrix size
gains_shape = (n_time, n_ant, n_chan, n_dir, 2)
# Jacobian shape
shape = gains_shape[1:]
# Covariance shape
covs_shape = (n_time, 2 * axis_length, 2 * axis_length)
# State vectors
m = np.zeros(gains_shape, dtype=np.complex128)
# Covariance matrices
P = np.zeros(covs_shape, dtype=np.complex128)
# Initial state and covariance
m[0] = gains_reshape(mp, shape)
P[0] = Pp
# Select CSR jacobian function
aug_jac = compute_aug_csr
# Calculate R^{-1} for a diagonal
Rinv = np.diag(1.0 / np.diag(R))
# Run Extended Kalman Filter with
# Sparse matrices
head = "==> Extended Kalman Filter (SPARSE): "
for k in range(1, n_time):
# Progress Bar
bar_len = 100
total = n_time - 1
filled_len = int(round(bar_len * k / float(n_time - 1)))
bar = u"\u2588" * filled_len + ' '\
* (bar_len - filled_len)
print("\r%s%d%%|%s| %d/%d" % (head, k / total * 100, bar, k, total),
end="")
# Predict Step
mp = gains_vector(m[k - 1])
Pp = P[k - 1] + Q
# Slice indices
start = tbin_indices[k - 1]
end = start + tbin_counts[k - 1]
# Calculate Slices
row_slice = slice(start, end)
vis_slice = vis[row_slice]
model_slice = model[row_slice]
weight_slice = weight[row_slice]
ant1_slice = ant1[row_slice]
ant2_slice = ant2[row_slice]
jones_slice = m[k - 1]
# Calculate Augmented Jacobian
J = aug_jac(model_slice, weight_slice, jones_slice, ant1_slice,
ant2_slice)
# Hermitian of Jacobian
J_herm = J.conjugate().T
# Calculate Measure Vector
y = measure_vector(vis_slice, weight_slice, n_ant, n_chan)
# Update Step
v = y - csr_dot_vec(J, mp)
Pinv = np.diag(1.0 / np.diag(Pp))
Tinv = np.linalg.inv(Pinv + J_herm @ Rinv @ J)
Sinv = Rinv - Rinv @ J @ Tinv @ J_herm @ Rinv
K = Pp @ J_herm @ Sinv
# Record Posterior values
m[k] = gains_reshape(mp + K @ v / 2.0, shape)
P[k] = np.diag(np.diag(Pp - K @ J @ Pp / 2.0).real)
# Newline
print()
# Return Posterior states and covariances
return m, P
def numba_algorithm(
mp : np.ndarray,
Pp : np.ndarray,
model : np.ndarray,
vis : np.ndarray,
weight : np.ndarray,
Q : np.ndarray,
R : np.ndarray,
ant1 : np.ndarray,
ant2 : np.ndarray,
tbin_indices : np.ndarray,
tbin_counts : np.ndarray,
alpha : np.float64):
"""Numpy-matrix implementation of EKF algorithm. It is
numba-compiled."""
# Time counts
n_time = len(tbin_indices)
# Number of Baselines
n_bl = model.shape[0]//n_time
# Number of Antennas
n_ant = int((np.sqrt(8*n_bl + 1) + 1)/2)
# Dimensions
n_chan, n_dir = model.shape[1], model.shape[2]
# Matrix-size
axis_length = n_ant * n_chan * n_dir
# Original matrix size
gains_shape = (n_time, n_ant, n_chan, n_dir, 2)
# Jacobian shape
shape = gains_shape[1:]
# Covariance shape
covs_shape = (n_time, 2*axis_length, 2*axis_length)
# State vectors
m = np.zeros(gains_shape, dtype=np.complex128)
# Covariance matrices
P = np.zeros(covs_shape, dtype=np.complex128)
# Initial state and covariance
m[0] = gains_reshape(mp, shape)
P[0] = Pp
# Select CSR jacobian function
aug_jac = compute_aug_np
# Calculate R^{-1} for a diagonal
Rinv = np.diag(1.0/np.diag(R))
I = np.eye(Rinv.shape[0])
# Run Extended Kalman Filter with
# NUMPY matrices
head = "==> Extended Kalman Filter (NUMPY|JIT): "
for k in range(1, n_time):
# Progress Bar in object-mode
with objmode():
progress_bar(head, n_time, k)
# Predict Step
mp = gains_vector(m[k - 1])
Pp = (P[k - 1] + Q)
# Slice indices
start = tbin_indices[k - 1]
end = start + tbin_counts[k - 1]
# Calculate Slices
row_slice = slice(start, end)
vis_slice = vis[row_slice]
model_slice = model[row_slice]
weight_slice = weight[row_slice]
ant1_slice = ant1[row_slice]
ant2_slice = ant2[row_slice]
jones_slice = m[k - 1]
# Calculate Augmented Jacobian
J = aug_jac(model_slice, weight_slice,
jones_slice, ant1_slice, ant2_slice)
# Hermitian of Jacobian
J_herm = J.conjugate().T
# Calculate Measure Vector
y = measure_vector(vis_slice, weight_slice,
n_ant, n_chan)
ym = measure_vector(model_slice[:, :, 0], weight_slice,
n_ant, n_chan)
# Update Step
#! NORMAL IMPLEMENTATION (FULL)
# v = y - J @ mp
# Pinv = np.diag(1.0/np.diag(Pp))
# Tinv = np.linalg.inv(Pinv + J_herm @ Rinv @ J)
# Sinv = Rinv - Rinv @ J @ Tinv @ J_herm @ Rinv
# K = Pp @ J_herm @ Sinv
# u = np.diag(J_herm @ J)
# z = J_herm @ v
# p = np.diag(Pp)
# pinv = 1.0/p
# mt = mp + alpha * p * (1.0 - np.diag(J_herm @ J)/(pinv + np.diag(J_herm @ J))) * (J_herm @ v)
# pt = p - alpha * p * (1.0 - u/(pinv + np.diag(J_herm @ J))) * np.diag(J_herm @ J) * p
# Record Posterior values
# m[k] = gains_reshape(mt, shape)
# P[k] = np.diag(pt.real)
# mt = mp + alpha * K @ v
# m[k] = gains_reshape(mp + alpha * K @ v, shape)
# P[k] = np.diag(np.diag(Pp - alpha * K @ J @ Pp).real)
# DIAGONALISATION IMPLEMENTATION
v = y - J @ mp
p = np.diag(Pp)
pinv = 1.0/p
u = diag_mat_dot_mat(J_herm, J) # Diagonal of JHJ
z = J_herm @ v # JHr
updt = alpha / (pinv + u)
est_m = mp + z * updt
est_P = (1 - alpha) * p + updt
# Record Posterior values
m[k] = gains_reshape(est_m, shape)
P[k] = np.diag(est_P.real)
# Newline
print()
# Return Posterior states and covariances
return m, P
def sparse_algorithm(mp: np.ndarray,
Pp: np.ndarray,
model: np.ndarray,
vis: np.ndarray,
weight: np.ndarray,
Q: np.ndarray,
R: np.ndarray,
ant1: np.ndarray,
ant2: np.ndarray,
tbin_indices: np.ndarray,
tbin_counts: np.ndarray,
tol: np.float64 = 1e-5,
maxiter: np.int32 = 20,
alpha: np.float64 = 0.5):
"""Sparse-matrix implementation of Iterated-EKF algorithm. Not
numba-compiled."""
# Time counts
n_time = len(tbin_indices)
# Number of Baselines
n_bl = model.shape[0] // n_time
# Number of Antennas
n_ant = int((np.sqrt(8 * n_bl + 1) + 1)) // 2
# Dimensions
n_chan, n_dir = model.shape[1:]
# Matrix-size
axis_length = n_ant * n_chan * n_dir
# Original matrix size
gains_shape = (n_time, n_ant, n_chan, n_dir, 2)
# Jacobian shape
shape = gains_shape[1:]
# Covariance shape
covs_shape = (n_time, 2 * axis_length, 2 * axis_length)
# State vectors
m = np.zeros(gains_shape, dtype=np.complex128)
# Covariance matrices
P = np.zeros(covs_shape, dtype=np.complex128)
# Initial state and covariance
m[0] = gains_reshape(mp, shape)
P[0] = Pp
# Select CSR jacobian function
aug_jac = compute_aug_csr
# Calculate R^{-1} for a diagonal
Rinv = np.diag(1.0 / np.diag(R))
# Run Iterated Extended Kalman Filter with
# Sparse matrices
head = "==> Iterated Extended Kalman Filter (SPARSE): "
for k in range(1, n_time):
# Progress Bar
progress_bar(head, n_time, k)
# Predict Step
mp = gains_vector(m[k - 1])
Pp = P[k - 1] + Q
# Slice indices
start = tbin_indices[k - 1]
end = start + tbin_counts[k - 1]
# Calculate Slices
row_slice = slice(start, end)
vis_slice = vis[row_slice]
model_slice = model[row_slice]
weight_slice = weight[row_slice]
ant1_slice = ant1[row_slice]
ant2_slice = ant2[row_slice]
# Current Iteration jones
jones_slice = gains_reshape(mp, shape)
# Calculate Augmented Jacobian
J = aug_jac(model_slice, weight_slice, jones_slice, ant1_slice,
ant2_slice)
# Hermitian of Jacobian
J_herm = J.conjugate().T
# Calculate Measure Vector
y = measure_vector(vis_slice, weight_slice, n_ant, n_chan)
# Initial state
mi = mp.copy()
# Current state
mn = mp
# Inverse of Prior Covariance
Pinv = np.diag(1.0 / np.diag(Pp))
# Iteration counter
i = 0
# State Estimation to reduce bias on
# estimation
while i < maxiter:
# Calculate Augmented Jacobian
J = aug_jac(model_slice, weight_slice, jones_slice, ant1_slice,
ant2_slice)
# Hermitian of Jacobian
J_herm = J.conjugate().T
# Update Step
v = y - csr_dot_vec(J, mi)
Tinv = np.linalg.inv(Pinv + J_herm @ Rinv @ J)
Sinv = Rinv - Rinv @ J @ Tinv @ J_herm @ Rinv
K = Pp @ J_herm @ Sinv
# Next state update
mt = mn + alpha * (mi - mn + K @ v)
# Stop if tolerance reached
if np.mean(np.abs(mt - mn)) <= tol:
break
# Else iterate next step
i += 1
mn = mt
# Next Iteration jones
jones_slice = gains_reshape(mn, shape)
# Record Posterior values from last iterated state
m[k] = gains_reshape(mt, shape)
P[k] = np.diag(np.diag(Pp - K @ J @ Pp).real)
# Newline
print()
# Return Posterior states and covariances
return m, P
def clean_meas_vec(data_slice, n_ant, n_chan):
clean_vis, _, _, weight, _, _, _ = data_slice
return measure_vector(clean_vis, weight, n_ant, n_chan)
