sampler = lambda nsamples : samplergaussian(mu,sigma2,alpha,nsamples)
#%%
folder_name = "example1"
try:
os.mkdir(folder_name)
except:
pass
nsamples = 5
samples = torch.linspace(-20,20,nsamples).reshape(-1,1)
#Just initiate with large mean and covariance
mu0 = torch.zeros(dim)
cov0 = (20.0/3)**2*torch.ones(dim)
#%% The main training class is initiated, it's gp model optimized,
# and initial component set
vb = VariationalBoosting(dim,logjoint,samples,mu0,cov0,
kernel_function="PMat",
matern_coef=2.5,
degree_hermite=60)
vb.optimize_bmc_model(maxiter=100,verbose=False)
#%% Tracking devices
vb.save_distribution("%s/distrib%i"%(folder_name,0))
nplot = 201
delta_x = torch.linspace(-20,20,nplot)
delta_x_np = delta_x.flatten().numpy()
tp_np = (logjoint(delta_x.reshape(-1,1)).cpu()).flatten().numpy()
prediction_np = (vb.bmcmodel.prediction(delta_x.reshape(-1,1),cov="none")*vb.evals_std + vb.evals_mean).\
numpy().astype(float)
dmeans = [torch.norm(vb.currentq_mean.to("cpu")-true_mean).numpy()]
dcovs = [torch.norm(vb.currentq_mean.to("cpu")-true_cov,2).numpy()]
elbo_list = [vb.evidence_lower_bound(nsamples=10000).cpu().numpy()]
step_list = [0]
time_list = [0]
torch.manual_seed(100) #For reproducibility
#Approximating nnormalized 2-d Cauchy
def logjoint(theta):
return torch.sum(-torch.log(1+theta**2),dim=-1)
#Set up parameters
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")
samples = sampling1(nsamples)
mu0 = torch.zeros(dim)
cov0 = 20.0 * torch.ones(dim)
#%%
#samples = vb.samples
training_interval = 20
acquisitions = ["prospective", "mmlt"]
vb = VariationalBoosting(dim,
logjoint,
samples,
mu0,
cov0,
bmc_type="FM",
normalization_mode="normalize",
training_space="gspace",
noise=1e-4,
kernel_function="PMat",
matern_coef=1.5,
numstab=-50.0,
degree_hermite=50)
vb.optimize_bmc_model(maxiter=500, verbose=1, lr=0.05)
#%%
elbo_list = [vb.evidence_lower_bound(nsamples=10000).cpu().numpy()]
kl_list = [vb.kullback_proposal_bmc(10000).item()]
step_list = [0]
time_list = [0.0]
#%%
print("Active sampling...")
try:
os.mkdir(folder_name)
except:
pass
nsamples = 51
ninducing = 50
samples = torch.linspace(-20, 20, nsamples).reshape(-1, 1)
#Just initiate with large mean and covariance
mu0 = torch.zeros(dim)
cov0 = (20.0 / 3)**2 * torch.ones(dim)
#%% The main training class is initiated, it's gp model optimized,
# and initial component set
vb = VariationalBoosting(dim,
logjoint,
samples,
mu0,
cov0,
kernel_function=kernels[0],
matern_coef=kernels[1],
degree_hermite=60)
vb.optimize_bmc_model(maxiter=100, verbose=False)
#%% Tracking devices
vb.save_distribution("%s/distrib%i" % (folder_name, 0))
nplot = 201
delta_x = torch.linspace(-20, 20, nplot)
delta_x_np = delta_x.flatten().numpy()
tp_np = (logjoint(delta_x.reshape(-1, 1)).cpu()).flatten().numpy()
prediction_np = (vb.bmcmodel.prediction(delta_x.reshape(-1,1),cov="none")*vb.evals_std + vb.evals_mean).\
numpy().astype(float)
dmeans = [torch.norm(vb.currentq_mean.to("cpu") - true_mean).numpy()]
dcovs = [torch.norm(vb.currentq_mean.to("cpu") - true_cov, 2).numpy()]
elbo_list = [vb.evidence_lower_bound(nsamples=10000).cpu().numpy()]