def condition_causal(self,
their_tensors,
absent_tensors,
movie_inds,
num_samples=1000):
their_cond = pyro.condition(model, data={"x": their_tensors})
absent_cond = pyro.condition(model, data={"x": absent_tensors})
their_y = []
for _ in range(num_samples):
their_y.append(
torch.sum(their_cond(
self.step2_params)['y'][movie_inds]).item())
absent_y = []
for _ in range(num_samples):
absent_y.append(
torch.sum(absent_cond(
self.step2_params)['y'][movie_inds]).item())
their_mean = np.mean(their_y)
absent_mean = np.mean(absent_y)
causal_effect_noconf = their_mean - absent_mean
return causal_effect_noconf
def loss_fn(design, num_particles, evaluation=False, **kwargs):
N, M = num_particles
expanded_design = lexpand(design, N)
# Sample from p(y, theta | d)
trace = poutine.trace(model).get_trace(expanded_design)
y_dict = {l: lexpand(trace.nodes[l]["value"], M) for l in observation_labels}
# Sample M times from q(theta | y, d) for each y
reexpanded_design = lexpand(expanded_design, M)
conditional_guide = pyro.condition(guide, data=y_dict)
guide_trace = poutine.trace(conditional_guide).get_trace(
y_dict, reexpanded_design, observation_labels, target_labels)
theta_y_dict = {l: guide_trace.nodes[l]["value"] for l in target_labels}
theta_y_dict.update(y_dict)
guide_trace.compute_log_prob()
# Re-run that through the model to compute the joint
modelp = pyro.condition(model, data=theta_y_dict)
model_trace = poutine.trace(modelp).get_trace(reexpanded_design)
model_trace.compute_log_prob()
terms = -sum(guide_trace.nodes[l]["log_prob"] for l in target_labels)
terms += sum(model_trace.nodes[l]["log_prob"] for l in target_labels)
terms += sum(model_trace.nodes[l]["log_prob"] for l in observation_labels)
terms = -terms.logsumexp(0) + math.log(M)
# At eval time, add p(y | theta, d) terms
if evaluation:
trace.compute_log_prob()
terms += sum(trace.nodes[l]["log_prob"] for l in observation_labels)
return _safe_mean_terms(terms)
def loss_fn(design, num_particles, evaluation=False, **kwargs):
expanded_design = lexpand(design, num_particles)
# Sample from p(y | d)
trace = poutine.trace(model).get_trace(expanded_design)
y_dict = {l: trace.nodes[l]["value"] for l in observation_labels}
theta_dict = {l: trace.nodes[l]["value"] for l in target_labels}
# Run through q(y | d)
qyd = pyro.condition(marginal_guide, data=y_dict)
marginal_trace = poutine.trace(qyd).get_trace(
expanded_design, observation_labels, target_labels)
marginal_trace.compute_log_prob()
# Run through q(y | theta, d)
qythetad = pyro.condition(likelihood_guide, data=y_dict)
cond_trace = poutine.trace(qythetad).get_trace(
theta_dict, expanded_design, observation_labels, target_labels)
cond_trace.compute_log_prob()
terms = -sum(marginal_trace.nodes[l]["log_prob"] for l in observation_labels)
# At evaluation time, use the right estimator, q(y | theta, d) - q(y | d)
# At training time, use -q(y | theta, d) - q(y | d) so gradients go the same way
if evaluation:
terms += sum(cond_trace.nodes[l]["log_prob"] for l in observation_labels)
else:
terms -= sum(cond_trace.nodes[l]["log_prob"] for l in observation_labels)
return _safe_mean_terms(terms)
def svi_model(self, x, thickness, intensity):
with pyro.plate('observations', x.shape[0]):
pyro.condition(self.model,
data={
'x': x,
'thickness': thickness,
'intensity': intensity
})()
def svi_model(self, x, age, sex, ventricle_volume, brain_volume):
with pyro.plate('observations', x.shape[0]):
pyro.condition(self.model,
data={
'x': x,
'sex': sex,
'age': age,
'ventricle_volume': ventricle_volume,
'brain_volume': brain_volume
})()
def svi_test():
rain_prob_prior = torch.tensor(.3)
my_sprinkler_prob_prior = torch.tensor(.6)
neighbor_sprinkler_prob_prior = torch.tensor(.2)
conditioned_lawn = pyro.condition(lawn,
data={
"my_lawn": torch.tensor([1.]),
"neighbor_lawn": torch.tensor([0.])
})
# guide = AutoGuide(lawn)
# set up the optimizer
adam_params = {"lr": 0.005, "betas": (0.90, 0.999)}
optimizer = Adam(adam_params)
# setup the inference algorithm
svi = SVI(conditioned_lawn, lawn_guide, optimizer, loss=Trace_ELBO())
n_steps = 1000
# do gradient steps
for step in range(n_steps):
svi.step(rain_prob_prior, my_sprinkler_prob_prior,
neighbor_sprinkler_prob_prior)
if step % 100 == 0:
print("step: ", step)
for p in [
'rain_prob', 'my_sprinkler_prob', 'neighbor_sprinkler_prob'
]:
print(p, ": ", pyro.param(p).item())
def importance_empirical_test():
conditioned_lawn = pyro.condition(lawn, data={"wet": torch.tensor([1.])})
rain_post = pyro.infer.Importance(conditioned_lawn, num_samples=100)
m = pyro.infer.EmpiricalMarginal(rain_post.run(),
sites=["rain", "sprinkler"])
print(m.log_prob(torch.tensor([1.])))
print(m.log_prob(torch.tensor([0.])))
def predict_cond(self, batch, par=None, **kwargs):
if par is "real":
par = self.get_real_par(batch)
else:
par = par(batch)
return pyro.condition(lambda batch: self.predict(batch, **kwargs),
data=par)(batch)
def likelihood(self, batch, par=None):
if par is None:
par = self.get_real_par(batch)
return pyro.condition(
lambda batch: self.forward(batch, mode="likelihood"),
data=par)(batch)
def step2_train(self, x_data, y_data, params, num_samples=1000):
print("Training Bayesian regression parameters...")
pyro.clear_param_store()
# Create a regression model
conditioned_on_x_and_y = pyro.condition(model,
data={
"x": x_data,
"y": y_data
})
svi = SVI(conditioned_on_x_and_y,
step2_guide,
self.step2_opt,
loss=Trace_ELBO(),
num_samples=num_samples)
for step in range(self.step2_iters):
loss = svi.step(params)
if step % 100 == 0:
print("[iteration %04d] loss: %.4f" %
(step + 1, loss / len(x_data)))
updated_params = {k: v for k, v in params.items()}
for name, value in pyro.get_param_store().items():
print("Updating value of hypermeter: {}".format(name))
updated_params[name] = value.detach()
print("Training complete.")
return updated_params
def step1_train(self, x_data, params):
conditioned_on_x = pyro.condition(model, data={"x": x_data})
svi = SVI(conditioned_on_x,
step1_guide,
self.step1_opt,
loss=Trace_ELBO())
print("\n Training Z marginal and W parameter marginal...")
# do gradient steps
pyro.get_param_store().clear()
for step in range(self.step1_iters):
loss = svi.step(params)
if step % 100 == 0:
print("[iteration %04d] loss: %.4f" %
(step + 1, loss / len(x_data)))
# grab the learned variational parameters
updated_params = {k: v for k, v in params.items()}
for name, value in pyro.get_param_store().items():
print("Updating value of hypermeter{}".format(name))
updated_params[name] = value.detach()
return updated_params
def test_undo_uncondition(self):
unconditioned_model = poutine.uncondition(self.model)
reconditioned_model = pyro.condition(
unconditioned_model, {"obs": torch.ones(2)}
)
reconditioned_trace = poutine.trace(reconditioned_model).get_trace()
assert_equal(reconditioned_trace.nodes["obs"]["value"], torch.ones(2))
def get_logprobs(self, **obs):
_required_data = ('x', 'thickness', 'intensity')
assert set(obs.keys()) == set(_required_data)
cond_model = pyro.condition(self.pyro_model.sample, data=obs)
model_trace = pyro.poutine.trace(cond_model).get_trace(
obs['x'].shape[0])
model_trace.compute_log_prob()
log_probs = {}
nats_per_dim = {}
for name, site in model_trace.nodes.items():
if site["type"] == "sample" and site["is_observed"]:
log_probs[name] = site["log_prob"].mean()
log_prob_shape = site["log_prob"].shape
value_shape = site["value"].shape
if len(log_prob_shape) < len(value_shape):
dims = np.prod(value_shape[len(log_prob_shape):])
else:
dims = 1.
nats_per_dim[name] = -site["log_prob"].mean() / dims
if self.hparams.validate:
print(
f'at site {name} with dim {dims} and nats: {nats_per_dim[name]} and logprob: {log_probs[name]}'
)
if torch.any(torch.isnan(nats_per_dim[name])):
raise ValueError('got nan')
return log_probs, nats_per_dim
def _get_pyro_model(
self,
posterior: bool = True,
num_observation: Optional[int] = None,
observation: Optional[torch.Tensor] = None,
) -> Callable:
"""Get model function for use with Pyro
If `num_observation` or `observation` is passed, the model is conditioned.
Args:
num_observation: Observation number
observation: Instead of passing an observation number, an observation may be
passed directly
posterior: If False, will mask prior which will result in model useful
for calculating log likelihoods instead of log posterior probabilities
"""
assert not (num_observation is not None and observation is not None)
if num_observation is not None:
observation = self.get_observation(num_observation=num_observation)
prior = self.get_prior()
simulator = self.get_simulator()
def model_fn():
prior_ = pyro.poutine.mask(prior, torch.tensor(posterior))
return simulator(prior_())
if observation is not None:
observation = self.unflatten_data(observation)
return pyro.condition(model_fn, {"data": observation})
else:
return model_fn
def loss_fn(design, num_particles, **kwargs):
try:
pyro.module("T", T)
except AssertionError:
pass
expanded_design = lexpand(design, num_particles)
# Unshuffled data
unshuffled_trace = poutine.trace(model).get_trace(expanded_design)
y_dict = {l: unshuffled_trace.nodes[l]["value"] for l in observation_labels}
# Shuffled data
# Not actually shuffling, resimulate for safety
conditional_model = pyro.condition(model, data=y_dict)
shuffled_trace = poutine.trace(conditional_model).get_trace(expanded_design)
T_joint = T(expanded_design, unshuffled_trace, observation_labels, target_labels)
T_independent = T(expanded_design, shuffled_trace, observation_labels, target_labels)
joint_expectation = T_joint.sum(0)/num_particles
A = T_independent - math.log(num_particles)
s, _ = torch.max(A, dim=0)
independent_expectation = s + ewma_log((A - s).exp().sum(dim=0), s)
loss = joint_expectation - independent_expectation
# Switch sign, sum over batch dimensions for scalar loss
agg_loss = -loss.sum()
return agg_loss, loss
def model_imp(num_iter, simulate):
conditioned_alive = pyro.condition(simulate.model_next_move_1,
data={"alive": 1})
imp = pyro.infer.Importance(conditioned_alive, num_samples=num_iter).run()
marginal_mv = pyro.infer.EmpiricalMarginal(imp, sites=["pl_mv"])
return marginal_mv()
def loss_fn(design, num_particles, evaluation=False, **kwargs):
expanded_design = lexpand(design, num_particles)
# Sample from p(y, theta | d)
trace = poutine.trace(model).get_trace(expanded_design)
y_dict = {l: trace.nodes[l]["value"] for l in observation_labels}
theta_dict = {l: trace.nodes[l]["value"] for l in target_labels}
# Run through q(theta | y, d)
conditional_guide = pyro.condition(guide, data=theta_dict)
cond_trace = poutine.trace(conditional_guide).get_trace(
y_dict, expanded_design, observation_labels, target_labels)
cond_trace.compute_log_prob()
if evaluation and analytic_entropy:
loss = mean_field_entropy(
guide,
[y_dict, expanded_design, observation_labels, target_labels],
whitelist=target_labels).sum(0) / num_particles
agg_loss = loss.sum()
else:
terms = -sum(cond_trace.nodes[l]["log_prob"]
for l in target_labels)
agg_loss, loss = _safe_mean_terms(terms)
return agg_loss, loss
def model(self, observations=None):
if isinstance(observations, dict):
data = observations[self._observation_name]
elif observations is not None:
data = observations
observations = {
self._observation_name:
observations.view(-1, *self._data_space)
}
else:
data = torch.zeros(1, *self._data_space)
observations = {}
data = data.view(data.shape[0], *self._data_space)
for module in self._category.children():
if isinstance(module, BaseModel):
module.set_batching(data)
min_depth = VAE_MIN_DEPTH if len(list(self.wiring_diagram)) == 1 else 0
morphism = self._category(self.wiring_diagram, min_depth=min_depth)
if observations is not None:
score_morphism = pyro.condition(morphism, data=observations)
else:
score_morphism = morphism
with pyro.plate('data', len(data)):
with name_pop(name_stack=self._random_variable_names):
output = score_morphism()
return morphism, output
def bvi(model, guide, claims, learning_rate=1e-5, num_samples=1):
"""perform blackbox mean field variational inference on simpleLCA.
This methods take a simpleLCA model as input and perform blackbox variational
inference, and returns a list of posterior distributions of hidden truth and source
reliability.
Concretely, if s is a source then posterior(s) is the probability of s being honest.
And if o is an object, or more correctly, is a random variable that has the support as
the domain of an object, then posterior(o) is the distribution over these support.
The underlying truth value of an object could be computed as the mode of this
distribution.
"""
data = make_observation_mapper(claims)
conditioned_lca = pyro.condition(lca_model, data=data)
pyro.clear_param_store() # is it needed?
svi = pyro.infer.SVI(model=conditioned_lca,
guide=lca_guide,
optim=pyro.optim.Adam({
"lr": learning_rate,
"betas": (0.90, 0.999)
}),
loss=pyro.infer.TraceGraph_ELBO(),
num_samples=num_samples)
return svi
def update_noise_importance(self, observed_steady_state, initial_noise):
observation_model = condition(self.noisy_model, observed_steady_state)
posterior = self.infer(observation_model, initial_noise)
updated_noise = {
k: EmpiricalMarginal(posterior, sites=k)
for k in initial_noise.keys()
}
return updated_noise
def get_conditioned_model(yaml_section, model, device="cpu"):
if yaml_section is None:
return model
conditions = {}
for name, val in yaml_section.items():
conditions[name] = yaml_params2._parse_val(name, val, device=device)
cond_model = pyro.condition(model, conditions)
return cond_model
def condition(self, condition_data: dict):
"""
It conditions self.model with condition data
Returns: Conditioned pyro model
"""
conditioned_model = pyro.condition(self.model, condition_data)
return conditioned_model
def condCausal(their_tensors, absent_tensors, movie_inds):
their_cond = pyro.condition(model, data={"x": their_tensors})
absent_cond = pyro.condition(model, data={"x": absent_tensors})
their_y = []
for _ in range(1000):
their_y.append(torch.sum(their_cond(p2)['y'][movie_inds]).item())
absent_y = []
for _ in range(1000):
absent_y.append(torch.sum(absent_cond(p2)['y'][movie_inds]).item())
their_mean = np.mean(their_y)
absent_mean = np.mean(absent_y)
causal_effect_noconf = their_mean - absent_mean
return causal_effect_noconf
def anwsers_given_traits(self, traits: np.ndarray):
# No need to do inference, as we condition on input variables
trait = torch.tensor(traits.reshape(1, self.t), dtype=torch.float32).expand(1_000, -1)
conditional_model = pyro.condition(self.model, data={
'trait': trait
})
return conditional_model(n_samples = 1_000)
def model_cond_sample(self, data_dict):
# sample the graph given the conditioned variables in data_dict
data_in = {}
for item in data_dict:
data_in[item] = data_dict[item]
cond_model = pyro.condition(self.model_sample, data=data_in)
return cond_model()
def observation(self, action):
action = action[0:len(action) - 1]
action_data = {
"action%d" % i: torch.tensor(float(action[i]))
for i in range(len(action))
}
print("Action: ", action)
#pyro importance sampling
action_cond = pyro.condition(self.model, data=action_data)
posterior = pyro.infer.Importance(action_cond, num_samples=800)
posterior = posterior.run()
marginal = posterior.marginal(
sites=['norm'] +
['reward_agent%d' % rew_no for rew_no in range(self.agent_no)])
empirical = marginal.empirical
#calculate norm posterior
inferred_norm = {}
for j in range(len(self.agent_norm)):
inferred_norm[j] = float(empirical['norm'].log_prob(j).exp())
#calculate reward posterior
inferred_reward = {}
for k in range(self.agent_no):
new_reward_dict = self.reward_dict.copy()
agent_reward = []
for rew in range(len(self.rew_prior['agent-0'])):
rew_possibility = float(empirical['reward_agent%d' %
k].log_prob(rew).exp())
new_reward_dict[rew] = [
i * rew_possibility for i in new_reward_dict[rew]
]
if rew == 0:
agent_reward = new_reward_dict[0]
else:
agent_reward = [
x + y
for x, y in zip(agent_reward, new_reward_dict[rew])
]
inferred_reward["agent-{}".format(k)] = agent_reward
#update norm prior
for i in range(len(self.agent_norm)):
self.n_prior[i] = empirical['norm'].log_prob(i).exp()
#update reward prior
for agent_no in range(self.agent_no):
for rew in range(len(self.rew_prior['agent-0'])):
self.rew_prior['agent-%d' % agent_no][rew] = empirical[
'reward_agent%d' % agent_no].log_prob(rew).exp()
print("Inferred norm: ", inferred_norm)
print("Inferred reward: ", inferred_reward)
return inferred_norm, inferred_reward
"""print("Observed Action: ")
def _get_log_likelihood(self, params):
"""Calculates the log-likelihood of the provided params.
Use `pyro.condition` and `pyro.poutine.trace`.
"""
# === Implement this
conditioned_model = pyro.condition(self.model, data=params)
trace = pyro.poutine.trace(conditioned_model).get_trace(
*self._model_args, **self._model_kwargs)
return trace.log_prob_sum()
def model(self, x: torch.Tensor):
"""
:param x: Tensor in shape (batch_size, x_dim)
:return:
"""
pyro.module('decoder', self.decoder)
batch_size = x.size()[0]
observed = pyro.condition(lambda: self.generate(batch_size),
data={'obs': x})()
return observed
def predict_cond(self, batch, par=None, **kwargs):
"""
Par can either be the string "real" or a function that outputs the parameters when called with a batch of data.
"""
if par is "real":
par = self.get_real_par(batch)
else:
par = par(batch)
return pyro.condition(lambda batch: self.predict(batch, **kwargs),
data=par)(batch)
def cond_predict(self, X_test, num_samples=1000):
if self.test_params is None:
self.infer_z(X_test)
cond = pyro.condition(model, data={"x": X_test})
predictions = np.zeros((X_test.shape[0]))
for _ in range(num_samples):
y_pred = cond(self.test_params)['y']
predictions += y_pred.detach().numpy()
return predictions / num_samples, self.test_params