def guide(x, y):
# parameters of (w : weight)
w_loc_1 = pyro.param('w_loc_1', torch.tensor(0.))
w_scale_1 = pyro.param('w_scale_1',
torch.tensor(1.),
constraint=constraints.positive)
# parameters of (b : bias)
b_loc_1 = pyro.param('b_loc_1', torch.tensor(0.))
b_scale_1 = pyro.param('b_scale_1',
torch.tensor(1.),
constraint=constraints.positive)
w_loc_2 = pyro.param('w_loc_2', torch.tensor(0.))
w_scale_2 = pyro.param('w_scale_2',
torch.tensor(1.),
constraint=constraints.positive)
b_loc_2 = pyro.param('b_loc_2', torch.tensor(0.))
b_scale_2 = pyro.param('b_scale_2',
torch.tensor(1.),
constraint=constraints.positive)
w1 = pyro.sample('w1', pdist.Normal(w_loc_1, w_scale_1))
b1 = pyro.sample('b1', pdist.Normal(b_loc_1, b_scale_1))
w2 = pyro.sample('w2', pdist.Normal(w_loc_2, w_scale_2))
b2 = pyro.sample('b2', pdist.Normal(b_loc_2, b_scale_2))
def model(x, y):
w = pyro.sample('w', pdist.Normal(0., 1.))
b = pyro.sample('b', pdist.Normal(0.5, 1.))
# define model
mean = w * x + b
# variance of distribution centered around y
# sigma = pyro.sample('sigma', pdist.Normal(0., 0.01))
pyro.sample('obs', pdist.Normal(mean, 0.01), obs=y)
return mean
def guide(data):
# define parameters
# loc and scale for latent variable `z_loc`
guide_loc = pyro.param('guide_loc', torch.tensor(0.))
guide_scale = pyro.param('guide_scale',
torch.tensor(1.),
constraint=constraints.positive)
# we would like to learn the distribution `loc`
pyro.sample('z_loc', pdist.Normal(guide_loc, guide_scale))
def guide(data):
""" Guide is a variational distribution """
# define parameters
# loc and scale for latent variable `z_loc`
guide_loc = pyro.param('guide_loc', torch.tensor(0.))
guide_scale = pyro.param('guide_scale', torch.tensor(0.)).exp()
# we would like to learn the distribution `loc`
pyro.sample('z_loc', pdist.Normal(guide_loc, guide_scale))
def model(x, y):
w = pyro.sample('w', pdist.Normal(torch.zeros(4), torch.ones(4)))
b = pyro.sample('b', pdist.Normal(0., 1.))
# define logistic regression model
y_hat = torch.sigmoid((w * x).sum(dim=1) + b)
# variance of distribution centered around y
# sigma = pyro.sample('sigma', pdist.Normal(0., 0.01))
pyro.sample('obs', pdist.Bernoulli(y_hat), obs=y)
def model(data):
""" Model : z_loc -> x
Args
data : list of scalars
"""
z_loc = pyro.sample('z_loc', pdist.Normal(0., 1.))
with pyro.plate('data', size=len(data), dim=-1):
# normally distributed observations
pyro.sample('obs', pdist.Normal(z_loc, 1.), obs=data)
def guide(x, y):
# parameters of (w : weight)
w_loc = pyro.param('w_loc', torch.zeros(4))
w_scale = pyro.param('w_scale',
torch.ones(4),
constraint=constraints.positive)
# parameters of (b : bias)
b_loc = pyro.param('b_loc', torch.tensor(0.))
b_scale = pyro.param('b_scale',
torch.tensor(1.),
constraint=constraints.positive)
w = pyro.sample('w', pdist.Normal(w_loc, w_scale))
b = pyro.sample('b', pdist.Normal(b_loc, b_scale))
def guide(x, y):
# parameters of (w : weight)
w_loc = pyro.param('w_loc', torch.tensor(1.))
w_scale = pyro.param('w_scale',
torch.tensor(1.),
constraint=constraints.positive)
# parameters of (b : bias)
b_loc = pyro.param('b_loc', torch.tensor(0.))
b_scale = pyro.param('b_scale',
torch.tensor(1.),
constraint=constraints.positive)
# parameters of (sigma)
# sigma_loc = pyro.param('sigma_loc', torch.tensor(0.), constraint=constraints.positive) # .exp()
# sample (w, b, sigma)
w = pyro.sample('w', pdist.Normal(w_loc, w_scale))
b = pyro.sample('b', pdist.Normal(b_loc, b_scale))
def model(x, y):
w1 = pyro.sample('w1', pdist.Normal(0., 1.))
b1 = pyro.sample('b1', pdist.Normal(3., 1.))
w2 = pyro.sample('w2', pdist.Normal(0., 1.))
b2 = pyro.sample('b2', pdist.Normal(3., 1.))
w3 = pyro.sample('w3', pdist.Normal(0., 1.))
b3 = pyro.sample('b3', pdist.Normal(3., 1.))
# define model
mean = w3 * torch.tanh(w2 * torch.tanh(w1 * x + b1) + b2) + b3
pyro.sample('obs', pdist.Normal(mean, 0.01), obs=y)
return mean
def model(data):
z_loc = pyro.sample('z_loc', pdist.Normal(0., 1.))
# normally distributed observations
# for datapoint in data:
pyro.sample('obs', pdist.Normal(z_loc, 1.), obs=data)