Ey = Wy @ h[1:] + by
Sig_y_diag = np.diag(Sigma_y).reshape(yd,1)
Sig_y_in = np.ones((T,yd,1))*Sig_y_diag
y= np.random.normal(Ey, np.sqrt(Sig_y_in))
'''
if train_weights == True:
#Update Weights
x = np.concatenate((h[:-1, :, :], u, np.ones((T, 1, 1))), axis=1)
xxT = (x[..., None] * x[:, None, :]).reshape(T, d + ud + 1, d + ud + 1)
Wz_bar, Wz, Uz, bz = update.Wbar_update(z, omega_z, x, xxT,
1 / Sigma_theta, Wz_mu_prior,
T, d, ud)
Wr_bar, Wr, Ur, br = update.Wbar_update(r, omega_r, x, xxT,
1 / Sigma_theta, Wr_mu_prior,
T, d, ud)
#slight adjustment to x for tanh approx
rx = 2 * np.concatenate(
(h[:-1, :, :] * r, u, np.ones((T, 1, 1))), axis=1)
rxrxT = (rx[..., None] * rx[:, None, :]).reshape(
T, d + ud + 1, d + ud + 1)
Wp_bar, Wp, Up, bp = update.Wbar_update(v, gamma, rx, rxrxT,
1 / Sigma_theta, Wp_mu_prior,
T, d, ud)
def gibbs_loop(N, N_burn, T, d,T_check, ud, yd, h0, inv_var, Sigma_y_inv,
Sigma_theta,
Sigma_y_theta,Sigma_y, Wz_mu_prior, Wr_mu_prior, Wp_mu_prior,
Wy_mu_prior, Wz, Uz, bz, Wr, Ur, br, Wp, Up, bp, Wy, by,
train_weights, u, y, h, r, rh, z, log_check):
###
h_plot_samples = np.zeros((N,d))
log_joint_vec = []
###
h_samples_vec = np.zeros((N-N_burn-1,T,d))
h_samples_vec2 = np.zeros((N-N_burn-1,T+1,d))
z_samples_vec = np.zeros((N-N_burn-1,T,d))
r_samples_vec = np.zeros((N-N_burn-1,T,d))
v_samples_vec = np.zeros((N-N_burn-1,T,d))
Wz_bar_samples_vec = np.zeros((N-N_burn-1,d,d+ud+1))
Wr_bar_samples_vec = np.zeros((N-N_burn-1,d,d+ud+1))
Wp_bar_samples_vec = np.zeros((N-N_burn-1,d,d+ud+1))
Wy_bar_samples_vec = np.zeros((N-N_burn-1,yd,d+1))
h_samples =0
z_samples =0
r_samples =0
v_samples =0
Wz_bar_samples = 0
Wr_bar_samples = 0
Wp_bar_samples = 0
Wy_bar_samples = 0
Wz_bar = Wz_mu_prior
Wr_bar = Wr_mu_prior
Wp_bar = Wp_mu_prior
Wy_bar = Wy_mu_prior
for k in range(0,N):
#Update pgs
omega_z = update.pg_update(1, h, u, Wz, Uz, bz, T, d)
omega_r = update.pg_update(1, h, u, Wr, Ur, br, T, d)
rh[:-1,:,:] = r*h[:-1,:,:]
gamma = update.pg_update(1, rh, u, 2*Wp, 2*Up, 2*bp, T, d)
#Update v
fv = build.build_v_param(h,z,rh,u,Wp,Up,bp, inv_var, d)
Ev = update.update_bern(fv)
v = np.random.binomial(1, Ev, size=(T,d,1))
#Update Zs
fz = build.build_z_param(h,v,inv_var.reshape(d,1), Wz, Uz, u, bz, d)
Ez = update.update_bern(fz)
z = np.random.binomial(1,Ez, size=(T,d,1))
#Update r's
for j in range(0,d):
frd = build.build_rd_param(h,u,v,gamma,r,Wp,Up,bp,Wr,Ur,br,j)
Erd = update.update_bern(frd)
r[:,j,:] = np.random.binomial(1,Erd, size=(T,1))
#Update hs
#Kalman sampling
J_dyn_11,J_dyn_22,J_dyn_21,J_obs,J_ini = build.build_prec_x_kalman(
inv_var, Sigma_y_inv, Wy,
Wz, omega_z, z, Wr, omega_r, r,
Wp, gamma, v, T, d)
h_dyn_1, h_dyn_2, h_obs, h_ini = build.build_prec_muT_kalman(
h0, u, inv_var, y, Sigma_y_inv, Wy, by,
z, omega_z, Wz, Uz, bz,
r, omega_r, Wr, Ur, br,
v, Wp, Up, bp, gamma, T, d)
log_Z_obs = np.zeros(T)
h =messages.kalman_info_sample(J_ini, h_ini, 0, J_dyn_11, J_dyn_21,
J_dyn_22, h_dyn_1, h_dyn_2, 0,
J_obs, h_obs, log_Z_obs)
h = h.reshape(T,d,1)
h = np.concatenate((h0.reshape(1,d,1), h), axis=0)
'''
#Full T*dXT*d matrix
prec = build.build_prec_x(inv_var, Sigma_y_inv, Wy,
Wz, omega_z, z, Wr, omega_r, r,
Wp, gamma, v, T, d)
prec_muT = build.build_prec_muT(h0, u, inv_var, y, Sigma_y_inv, Wy, by,
z, omega_z, Wz, Uz, bz,
r, omega_r, Wr, Ur, br,
v, Wp, Up, bp, gamma, T, d)
mu, covar = update.update_normal_dim(prec,prec_muT)
h = np.random.multivariate_normal(mu[:,0],covar)
h = h.reshape(T,d,1)
h = np.concatenate((h0.reshape(1,d,1), h), axis=0)
'''
'''
#Sample y's, for Testing purposes
Ey = Wy @ h[1:] + by
Sig_y_diag = np.diag(Sigma_y).reshape(yd,1)
Sig_y_in = np.ones((T,yd,1))*Sig_y_diag
y= np.random.normal(Ey, np.sqrt(Sig_y_in))
'''
if train_weights == True:
#Update Weights
x = np.concatenate((h[:-1,:,:],u, np.ones((T,1,1))), axis=1)
xxT = (x[...,None]*x[:,None,:]).reshape(T,d+ud+1,d+ud+1)
Wz_bar, Wz, Uz, bz = update.Wbar_update(z,omega_z,x,xxT,
1/Sigma_theta,
Wz_mu_prior,T,d,ud)
Wr_bar, Wr, Ur, br = update.Wbar_update(r,omega_r,x,xxT,
1/Sigma_theta,
Wr_mu_prior,T,d,ud)
#slight adjustment to x for tanh approx
rx = 2*np.concatenate((h[:-1,:,:]*r,u, np.ones((T,1,1))), axis=1)
rxrxT = (rx[...,None]*rx[:,None,:]).reshape(T,d+ud+1,d+ud+1)
Wp_bar, Wp, Up, bp = update.Wbar_update(v, gamma, rx, rxrxT,
1/Sigma_theta,
Wp_mu_prior,T,d,ud)
#Update y weights
x = np.concatenate((h[1:,:,:], np.ones((T,1,1))), axis=1)
xxT = (x[...,None]*x[:,None,:]).reshape(T,d+1,d+1)
Wy_bar, Wy, by = update.Wbar_y_update(x,xxT,y, Sigma_y_inv,
1/Sigma_y_theta, Wy_mu_prior,
T,d,yd)
if k > N_burn:
h_samples_vec[k-N_burn-1] = h[1:].reshape(T,d)
h_samples_vec2[k-N_burn-1] = h.reshape(T+1,d)
z_samples_vec[k-N_burn-1] = z.reshape(T,d)
r_samples_vec[k-N_burn-1] = r.reshape(T,d)
v_samples_vec[k-N_burn-1] = v.reshape(T,d)
h_samples += h
z_samples += z
r_samples += r
v_samples += v
Wz_bar_samples += Wz_bar
Wz_bar_samples_vec[k-N_burn-1] = Wz_bar
Wr_bar_samples += Wr_bar
Wr_bar_samples_vec[k-N_burn-1] = Wr_bar
Wp_bar_samples += Wp_bar
Wp_bar_samples_vec[k-N_burn-1] = Wp_bar
Wy_bar_samples += Wy_bar
Wy_bar_samples_vec[k-N_burn-1] = Wy_bar
print(k)
###
h_plot_samples[k] = h[T_check,:,0]
sig_h_inv = np.diag(inv_var)
if k%log_check == 0:
log_joint_vec.append( log_prob.marg_full_log_joint_no_weights(T, d, yd,
u, y, h, sig_h_inv, z, v, r,
Wz, Uz, bz, Wr, Ur, br,
Wp, Up, bp, Wy, by, Sigma_y_inv) )
###
return h_samples, z_samples, r_samples, v_samples, Wz_bar_samples, Wr_bar_samples, Wp_bar_samples, Wy_bar_samples, h_samples_vec, Wz_bar_samples_vec, Wr_bar_samples_vec, Wp_bar_samples_vec, Wy_bar_samples_vec, h_plot_samples, log_joint_vec, z_samples_vec, r_samples_vec, v_samples_vec, h_samples_vec2
#Update hs
prec = build.build_prec_x(inv_var, W, omega, z, T, d)
prec_muT = build.build_prec_muT(h0, u, inv_var, z, omega, W, U, b, T, d)
mu, covar = update.update_normal_dim(prec, prec_muT)
h = np.random.multivariate_normal(mu[:, 0], covar)
h = h.reshape(T, d, 1)
h = np.concatenate((h0.reshape(1, d, 1), h), axis=0)
if train_weights == True:
#Update Weights
x = np.concatenate((h[1:, :, :], u, np.ones((T, 1, 1))), axis=1)
xxT = (x[..., None] * x[:, None, :]).reshape(T, d + ud + 1, d + ud + 1)
W_covar, W_mu = update.Wbar_update(z, omega, x, xxT, 1 / Sigma_theta,
W_mu_prior, T, d, ud)
W_bar = update.sample_weights(W_covar, W_mu, d, ud)
W, U, b = update.extract_W_weights(W_bar, d, ud)
'''
if k%100 == 0:
loglike = build.log_like(h0,inv_var,h, z, omega, u, W, U, b, bpg,T, d)
print(loglike)
log_like_vec.append(loglike[0,0,0])
'''
if k > N_burn:
h_samples += h
z_samples += z
W_bar_samples += W_bar
print(k)
Eh = h_samples / (N - N_burn - 1)