def comp(theta, xx):
resid_old = []
resid_new = []
A_old = np.zeros([T, M, NN], dtype=complex)
A_new = np.zeros([T, M, NN], dtype=complex)
theta_matrix = np.diag(theta)
for t in range(T):
# A[t] = np.dot(np.dot(G[t], theta_matrix), Hr[t]) + Hd[t]
A_old[t] = Hd[t]
A_new[t] = np.dot(np.dot(G[t], theta_matrix), Hr[t]) + Hd[t]
# print(A_new[t].shape)
y_old = np.zeros([T, M, 1], dtype=complex)
y_new = np.zeros([T, M, 1], dtype=complex)
for t in range(T):
y_old[t] = np.dot(A_old[t], xx[t])
y_new[t] = np.dot(A_new[t], xx[t])
x_array_old, resid_array_old = pa.PIA_ASP(y_old, A_old, sp=0)
x_array_new, resid_array_new = pa.PIA_ASP(y_new, A_new, sp=0)
print('resid_old', resid_array_old[-1] / np.linalg.norm(A_old[t]))
print('resid_new', resid_array_new[-1] / np.linalg.norm(A_new[t]))
return resid_array_old[-1] / np.linalg.norm(
A_old[t]), resid_array_new[-1] / np.linalg.norm(A_new[t])
def comp(theta, xx, num):
A_old = np.zeros([T, M, NN], dtype=complex)
A_new = np.zeros([T, M, NN], dtype=complex)
theta_matrix = np.diag(theta)
for t in range(T):
# A[t] = np.dot(np.dot(G[t], theta_matrix), Hr[t]) + Hd[t]
A_old[t] = Hd[t]
# 这个应该是由theta_optimization 来决定的 A_new[t] = np.dot(np.dot(G[t], theta_matrix), Hr[t]) + Hd[t]
A_new[t] = np.dot(np.dot(G[t, :, 0:num], theta_matrix),
Hr[t, 0:num, :]) + Hd[t]
y_old = np.zeros([T, M, 1], dtype=complex)
y_new = np.zeros([T, M, 1], dtype=complex)
for t in range(T):
y_old[t] = np.dot(A_old[t], xx[t])
y_new[t] = np.dot(A_new[t], xx[t])
before = time.time()
x_array_old, resid_array_old = pa.PIA_ASP(y_old, A_old, sp=50)
after = time.time()
print('old', after - before)
before = time.time()
x_array_new, resid_array_new = pa.PIA_ASP(y_new, A_new, sp=50)
after = time.time()
print('new', after - before)
resid_old = resid_array_old[-1]
resid_new = resid_array_new[-1]
# print('resid_old', resid_array_old[-1] / np.linalg.norm(A_old[t]))
# print('resid_new', resid_array_new[-1] / np.linalg.norm(A_new[t]))
# print('resid_old', resid_array_old[-1] / np.linalg.norm(y_old[t]))
# print('resid_new', resid_array_new[-1] / np.linalg.norm(y_new[t]))
return resid_array_old[-1] / np.linalg.norm(
y_old[t]), resid_array_new[-1] / np.linalg.norm(y_new[t])
def comp(theta, xx):
A_old = np.zeros([T, M, NN], dtype=complex)
A_new = np.zeros([T, M, NN], dtype=complex)
theta_matrix = np.diag(theta)
for t in range(T):
# A[t] = np.dot(np.dot(G[t], theta_matrix), Hr[t]) + Hd[t]
A_old[t] = Hd[t]
# 这个应该是由theta_optimization 来决定的 A_new[t] = np.dot(np.dot(G[t], theta_matrix), Hr[t]) + Hd[t]
A_new[t] = np.dot(np.dot(G[t], theta_matrix), Hr[t]) + Hd[t]
# print(A_new[t].shape)
y_old = np.zeros([T, M, 1], dtype=complex)
y_new = np.zeros([T, M, 1], dtype=complex)
for t in range(T):
y_old[t] = np.dot(A_old[t], xx[t])
y_new[t] = np.dot(A_new[t], xx[t])
x_array_old, resid_array_old = pa.PIA_ASP(y_old, A_old, sp=0)
x_array_new, resid_array_new = pa.PIA_ASP(y_new, A_new, sp=0)
resid_old = resid_array_old[-1]
resid_new = resid_array_new[-1]
print('resid_old', resid_array_old[-1] / np.linalg.norm(A_old[t]))
print('resid_new',
resid_array_new[-1] / np.linalg.norm(A_new[t]),
end='\t')
def optimize(n, x, t, the, theta_w, theta_record):
min_sum = 999
min_theta = 0
theta_matrix = np.zeros([N, N], dtype=complex)
for i in range(N):
theta_matrix[i][i] = math.cos(the[i]) + 1j * math.sin(the[i])
for th in theta_array:
theta_matrix[n][n] = th
A = np.zeros([T, M, NN], dtype=complex)
for t in range(T):
A[t] = np.dot(np.dot(G[t], theta_matrix), Hr[t]) + Hd[t]
# print(A.shape)
y = np.zeros([T, M, 1], dtype=complex)
for t in range(T):
y[t] = np.dot(A[t], (x[t] + sigma[t]))
np.savetxt('/Users/xfn/Downloads/cvx/projects/data/sigma.csv',
sigma[0],
delimiter=',')
# print(y.shape)
x_array, resid_array = pa.PIA_ASP(y, A, sp=0)
# print(resid_array)
sum_else = XL[n] * abs(th - theta_w[n]) + (
rho + L) / 2 * np.linalg.norm(th - theta_w[n])
# sum_else = XL[n] * (th - theta_w[n]) + (rho + L) / 2 * np.linalg.norm(th - theta_w[n])
# sum_else = 0
# print(resid_array[-1])
# print('sum_else', sum_else)
if resid_array[-1] + sum_else < min_sum:
# if sum_else < min_sum:
min_sum = resid_array[-1] + sum_else
# min_sum = sum_else
min_theta = th
# A = np.dot(np.dot(G, np.diag(theta)), Hr) + Hd
# sumthetan = 0
# for i in range(M):
# sumthetan += derivativeTheta(i, n, x, t) * derivativeTheta(i, n, x, t)
#
# sumthetan += XL[n] * (theta_wan - theta) + (rho + L) / 2 * np.linalg.norm(theta_wan - theta)
return min_theta
def optimize(n, x, t, the, theta_w, theta_record, num):
min_sum = 999
min_theta = 0
theta_matrix = np.zeros([num, num], dtype=complex)
for i in range(num):
theta_matrix[i][i] = math.cos(the[i]) + 1j * math.sin(the[i])
for th in theta_array:
theta_matrix[n][n] = th
A = np.zeros([T, M, NN], dtype=complex)
for t in range(T):
A[t] = np.dot(np.dot(G[t, :, 0:num], theta_matrix),
Hr[t, 0:num, :]) + Hd[t] # print(A.shape)
y = np.zeros([T, M, 1], dtype=complex)
for t in range(T):
y[t] = np.dot(A[t], x[t])
# print(y.shape)
x_array, resid_array = pa.PIA_ASP(y, A, sp=50)
# print(resid_array)
sum_else = XL[n] * abs(th - theta_w[n]) + (
rho + L) / 2 * np.linalg.norm(th - theta_w[n])
# sum_else = XL[n] * (th - theta_w[n]) + (rho + L) / 2 * np.linalg.norm(th - theta_w[n])
# sum_else = 0
# print(resid_array[-1])
# print('sum_else', sum_else)
if resid_array[-1] + sum_else < min_sum:
# if sum_else < min_sum:
min_sum = resid_array[-1] + sum_else
# min_sum = sum_else
min_theta = th
# A = np.dot(np.dot(G, np.diag(theta)), Hr) + Hd
# sumthetan = 0
# for i in range(M):
# sumthetan += derivativeTheta(i, n, x, t) * derivativeTheta(i, n, x, t)
#
# sumthetan += XL[n] * (theta_wan - theta) + (rho + L) / 2 * np.linalg.norm(theta_wan - theta)
return min_theta
KK = 20
Hr = np.random.uniform(low=1e-6, high=1e-4, size=[T, N, NN])
Hd = np.random.uniform(low=1e-13, high=1e-10, size=[T, M, NN])
G = np.random.uniform(low=1e-8, high=1e-6, size=[T, M, N])
x = np.load('x.npy')
theta_opt = np.load('theta.npy')
theta_matrix = np.diag(theta_opt)
if __name__ == "__main__":
# x = dataset.setRelativity(NN, T, K, KK)
A_old = np.zeros([T, M, NN], dtype=complex)
A_new = np.zeros([T, M, NN], dtype=complex)
for t in range(T):
# A[t] = np.dot(np.dot(G[t], theta_matrix), Hr[t]) + Hd[t]
A_old[t] = Hd[t]
#这个应该是由theta_optimization 来决定的 A_new[t] = np.dot(np.dot(G[t], theta_matrix), Hr[t]) + Hd[t]
A_new[t] = np.dot(np.dot(G[t], theta_matrix), Hr[t]) + Hd[t]
y_old = np.zeros([T, M, 1], dtype=complex)
y_new = np.zeros([T, M, 1], dtype=complex)
for t in range(T):
y_old[t] = np.dot(A_old[t], x[t])
y_new[t] = np.dot(A_new[t], x[t])
x_array_old, resid_array_old = pa.PIA_ASP(y_old, A_old, sp=50)
x_array_new, resid_array_new = pa.PIA_ASP(y_new, A_new, sp=50)
print('resid_old', resid_array_old[-1])
print('resid_new', resid_array_new[-1])