#mean_grad_rms_analytical = log_p_grad_rms_analytical[:K]
#log_sigma_grad_rms_analytical = log_p_grad_rms_analytical[K:]
#mean_grad_clr2_analytical = log_p_grad_clr2_analytical[:K]
#log_sigma_grad_clr2_analytical = log_p_grad_clr2_analytical[K:]
#mean_grad_clr_analytical = log_p_grad_clr_analytical[:K]
#log_sigma_grad_clr_analytical = log_p_grad_clr_analytical[K:]
#log_p_grad = log_p_grad_analytical
#log_p_grad_clr = log_p_grad_clr_analytical
#log_p_grad_rms = log_p_grad_rms_analytical
#log_p_grad_clr2 = log_p_grad_clr2_analytical
log_p = np.array([fit_hmc.log_prob(s) for s in samples.T])
elbo = np.mean(log_p) + compute_entropy(log_sigmas)
#elbo_analytical = elbo_logit(zs, means, log_sigmas, X, Y)
log_p_clr = np.array(
[fit_hmc.grad_log_prob(s) for s in samples_clr.T])
elbo_clr = np.mean(log_p_clr) + compute_entropy(log_sigmas_vb_clr)
#elbo_clr_analytical = elbo_logit(zs, means_vb_clr, log_sigmas_vb_clr, X, Y)
log_p_clr2 = np.array(
[fit_hmc.grad_log_prob(s) for s in samples_clr2.T])
elbo_clr2 = np.mean(log_p_clr2) + compute_entropy(
log_sigmas_vb_clr2)
#elbo_clr2_analytical = elbo_logit(zs, means_vb_clr2, log_sigmas_vb_clr2, X, Y)
log_p_rms = np.array(
[fit_hmc.grad_log_prob(s) for s in samples_rms.T])
zs_t = np.random.standard_t(dof, size=(num_params, num_samples))
zs_t_swa = np.random.standard_t(dof, size=(num_params, num_samples_swa))
samples, grad_correction = reparametrize(zs, means, L)
samples_rms, grad_correction_rms = reparametrize(zs, means_vb_rms, L_vb_rms)
samples_clr, grad_correction_clr = reparametrize(zs, means_vb_clr, L_vb_clr)
#samples_swa2, grad_correction_swa2 = reparametrize(zs, means_vb_clr2, log_sigmas_vb_clr2)
#samples, grad_correction = reparametrize(zs_swa, means, log_sigmas)
#log_p_grad1 = np.array([fit_hmc.grad_log_prob(s)])
# evaluate gradient of log p (does grad_log_prob support vectorization??) and gradient of log q
log_p_grad = np.array([fit_hmc.grad_log_prob(s) for s in samples.T])
log_p_grad_rms = np.array([fit_hmc.grad_log_prob(s) for s in samples_rms.T])
log_p_grad_clr = np.array([fit_hmc.grad_log_prob(s) for s in samples_clr.T])
log_p = np.array([fit_hmc.log_prob(s) for s in samples.T])
elbo= np.mean(log_p) + compute_entropy(L)
log_p_clr = np.array([fit_hmc.grad_log_prob(s) for s in samples_clr.T])
elbo_clr = np.mean(log_p_clr) + compute_entropy(L_vb_clr)
log_p_rms = np.array([fit_hmc.grad_log_prob(s) for s in samples_rms.T])
elbo_rms = np.mean(log_p_rms) + compute_entropy(L_vb_rms)
entropy_grad_swa = np.diag(1./np.diag(L_vb_swa))
entropy_grad_clr = np.diag(1./np.diag(L_vb_clr))
entropy_grad_rms = np.diag(1./np.diag(L_vb_rms))
#print(log_p_grad)
#print( np.mean(log_p_grad*grad_correction, axis=0) + entropy_grad)
#print(log_p_grad_analytical )
# compute gradients wrt. mean and log_sigma
mean_grad = np.mean(log_p_grad, axis=0)
mean_grad_clr = np.mean(log_p_grad_clr, axis=0)