Implementation of two sparse Gaussian process algorithms:
- VFE
- SVI-GP
See demo.py
from VFE import VFE
from math import pi
import torch
import numpy as np
torch.set_default_tensor_type('torch.DoubleTensor')
train_x = torch.tensor(np.loadtxt('./datasets/kin_40k/train_x'))
train_y = torch.tensor(np.loadtxt('./datasets/kin_40k/train_y'))
test_x = torch.tensor(np.loadtxt('./datasets/kin_40k/test_x'))
test_y = torch.tensor(np.loadtxt('./datasets/kin_40k/test_y'))
conf = dict()
conf['num_inducing'] = 256
conf['num_epoch'] = 300
conf['jitter_u'] = 1e-6
conf['kmeans'] = True
model = VFE(train_x, train_y, conf)
model.train_scipy()
py, ps2 = model.predict(test_x)
mse = torch.mean((test_y - py)**2)
smse = mse / torch.var(test_y)
log_prob1 = -0.5 * torch.log(2 * pi * ps2) - 0.5 * (test_y - py)**2 / ps2
log_prob2 = -0.5 * torch.log(2 * pi * train_y.var()) - 0.5 * (test_y - train_y.mean())**2 / train_y.var()
snlp = -1 * (log_prob1 - log_prob2).mean()
print('smse = %g' % smse)
print('snlp = %g' % snlp)
np.savetxt('py', py.detach().numpy())
np.savetxt('ps2', ps2.detach().numpy())
np.savetxt('test_y', test_y.detach().numpy())