def get_actionStepLength(self) -> float:
actionStepLength = 0.0
if 17 <= self.age < 45:
actionStepLength = distrs.Uniform(
-0.1, 0.1)().item() # lower_bound = -0.1, upper_bound = 0.1
else:
actionStepLength = distrs.Chi2(2)().item(
) / 16 - 0.025 # k = 2 -> mean = 0.125, sigma = 0.125
return actionStepLength
def guide():
# variational parameters
var_mu = pyro.param('var_mu', torch.tensor(180.))
var_mu_sig = pyro.param('var_mu_sig',
torch.tensor(5.),
constraint=constraints.positive)
var_sig = pyro.param('var_sig', torch.tensor(5.))
# factorized distribution
pyro.sample('mu', dist.Normal(loc=var_mu, scale=var_mu_sig))
pyro.sample('sigma', dist.Chi2(var_sig))
def get_sigma(self) -> float:
sigma = 0.0
if 17 <= self.age < 45:
sigma = random_noise().item()
elif 45 <= self.age < 55:
sigma = distrs.Normal(0.1,
0.05)().item() # mean = 0.1, sigma = 0.05
else:
sigma = distrs.Chi2(
2)().item() / 10 # k = 2 -> mean = 0.1 , sigma = 0.2
return sigma
def _params2probs(self, params):
"""Convert parameters to probability distributions.
Args:
params (dict[str -> np.array]): dictionary of distribution parameters
Returns:
dict[str -> Dist]: dictionary of prior probabilities
"""
M = dist.Normal(params["M_mu"],
self.softplus(params["M_sigma"])).independent(2)
beta = dist.Beta(self.softplus(params["beta_p"]),
self.softplus(params["beta_q"])).independent(1)
z_epsilon = dist.Beta(
self.softplus(params["z_epsilon_p"]),
self.softplus(params["z_epsilon_q"]))
scale_factor = dist.Chi2(params["scale_factor"])
return {"M": M, "beta": beta, "z_epsilon": z_epsilon,
"scale_factor": scale_factor}
def ChiSq(_name, df):
return {'x': pyro.sample(_name, dist.Chi2(df))}
pyro.clear_param_store()
pyro.enable_validation(True)
svi = SVI(model,
guide,
optim=pyro.optim.ClippedAdam({"lr": 0.01}),
loss=Trace_ELBO())
# do gradient steps
c = 0
for step in range(1000):
c += 1
loss = svi.step()
if step % 100 == 0:
print("[iteration {:>4}] loss: {:.4f}".format(c, loss))
sigma = dist.Chi2(pyro.param('var_sig')).sample((10000, )).numpy()
mu = dist.Normal(pyro.param('var_mu'), pyro.param('var_mu_sig')).sample(
(10000, )).numpy()
plt.figure()
plt.scatter(mu, sigma, alpha=0.01)
plt.xlim([extent[0], extent[1]])
plt.ylim([extent[2], extent[3]])
plt.ylabel('$\sigma$')
plt.xlabel('$\mu$')
plt.title('VI samples')
save_fig('bayes_unigauss_2d_pyro_post.pdf')
plt.show()