for i in range(100):
tictoc.tic()
_ = vb.update(maxiter_nc=300,
lr_nc=0.01,
n_samples_nc=500,
n_samples_sn=300,
n_iter_sn=300)
if ((i + 1) % 10) == 0:
try:
vb.update_bmcmodel(acquisition=acquisition, vreg=1e-3)
except:
continue
samples_dict[i + 1] = (vb.samples.numpy(), vb.evaluations.numpy())
# vb.update_full()
vb.cutweights(1e-3)
#%% Save trackings
elapsed = tictoc.toc(printing=False)
dmeans.append(np.linalg.norm(vb.currentq_mean.cpu() - true_mean, 2))
dcovs.append(np.linalg.norm(vb.currentq_cov.cpu() - true_cov, 2))
elbo_list.append(vb.evidence_lower_bound(nsamples=10000).cpu().numpy())
step_list.append(i + 1)
time_list.append(elapsed)
vbp = vb.current_logq(delta_x.reshape(
-1, 1)).cpu().flatten().numpy().astype(float)
vbp_list.append(vbp)
prediction_np = (vb.bmcmodel.prediction(delta_x.reshape(-1,1),cov="none")*vb.evals_std + vb.evals_mean).\
numpy().astype(float).flatten()
prediction_list.append(prediction_np)
# vb.save_distribution("%s/distrib%i"%(folder_name,i+1))
print(vb_utils.kl_vb_bmc(vb, 1000))
dim=2 #Dimension of problem
samples = torch.randn(20,dim) #Initial samples
mu0 = torch.zeros(dim) #Initial mean
cov0 = 20.0*torch.ones(dim) #Initial covariance
acquisition = "prospective" #Acquisition function
#Initialize algorithm
vb = VariationalBoosting(dim,logjoint,samples,mu0,cov0)
vb.optimize_bmc_model() #Optimize GP model
vb.update_full() #Fit first component
#Training loop
for i in range(100):
_ = vb.update() #Choose new boosting component
vb.update_bmcmodel(acquisition=acquisition) #Choose new evaluation
vb.cutweights(1e-3) #Weights prunning
if ((i+1)%20) == 0:
vb.update_full(cutoff=1e-3) #Joint parameter updating
vb.save_distribution("finaldistrib") #Save distribution
#%%
import math
distrib = torch.load("finaldistrib")
nplot = 21
x,y = torch.linspace(-6,6,nplot),torch.linspace(-6,6,nplot)
X,Y = torch.meshgrid(x,y)
Z1 = logjoint(torch.stack([X,Y],dim=-1).reshape(-1,2)).reshape(*X.shape)-\
2*math.log(math.pi)
Z2 = distrib.logprob(torch.stack([X,Y],dim=-1).reshape(-1,2)).reshape(*X.shape)
#%%
import matplotlib.pyplot as plt