def get_recon_error(recon, x, sigma):
if x.shape[1] == 1: # Binary image
ll = Bernoulli(recon).log_prob(x)
elif x.shape[1] == 3: # RGB image
ll = Normal(recon, sigma).log_prob(x)
else:
NotImplementedError('X must be either 1 or 3')
return -ll.sum()
def generate_gg_blocks(N):
"""
Generate 'clean' data points using the ggblocks feature set
@param N: the number of data points to generate
This function doesn't sample points from the IBP: it uses a Bernoulli
distribution for Z
"""
import numpy as np
# we have to make sure there's at least one feature in each sample
Z = Bernoulli(0.5).sample((N, 4))
while (Z.sum(1) == 0).any():
msk = (Z.sum(1) == 0)
Z[msk] = Bernoulli(0.5).sample((msk.sum().item(), 4))
unique, counts = np.unique(Z.sum(1).numpy(), return_counts=True)
print("# of points with 1 feature", float(counts[0])/float(N))
A = gg_blocks()
return Z @ A
def sample_Z(N):
Z = Bernoulli(0.25).sample((N, 4))
while (Z.sum(1) == 0).any():
msk = (Z.sum(1) == 0)
Z[msk] = Bernoulli(0.25).sample((msk.sum().item(), 4))
return Z
def log_likelihood(X, y, w, use_cuda=False):
data = torch.as_tensor(X.values)
target = torch.as_tensor(y.values)
logits = (data * w).sum(1)
logp = Bernoulli(logits=logits).log_prob(target)
return logp.sum()
