beanmachine.ppl.model.statistical_model module

class beanmachine.ppl.model.statistical_model.StatisticalModel

Bases: object

Parent class to all statistical models implemented in Bean Machine.

Every random variable in the model needs to be defined with function declaration wrapped the bm.random_variable .

Every deterministic functional that a user would like to query during inference should be wrapped in a bm.functional .

Every parameter of the guide distribution that is to be learned via variational inference should be wrapped in a bm.param .

static functional(f: Callable[[beanmachine.ppl.model.statistical_model.P], torch.Tensor]) → Callable[[beanmachine.ppl.model.statistical_model.P], Union[beanmachine.ppl.model.rv_identifier.RVIdentifier, torch.Tensor]]

Decorator to be used for every query defined in statistical model, which are functions of bm.random_variable

@bm.random_variable
def foo():
  return Normal(0., 1.)

@bm.functional():
def bar():
  return foo() * 2.0

static get_func_key(wrapper, arguments) → beanmachine.ppl.model.rv_identifier.RVIdentifier

Creates a key to uniquely identify the Random Variable.

Parameters

• wrapper – reference to the wrapper function

• arguments – function arguments

Returns

Tuple of function and arguments which is to be used to identify a particular function call.

static param(init_fn)

Decorator to be used for params (variable to be optimized with VI).:

@bm.param
def mu():
  return torch.zeros(2)

@bm.random_variable
def foo():
  return Normal(mu(), 1.)

static random_variable(f: Callable[[beanmachine.ppl.model.statistical_model.P], torch.distributions.distribution.Distribution]) → Callable[[beanmachine.ppl.model.statistical_model.P], Union[beanmachine.ppl.model.rv_identifier.RVIdentifier, torch.Tensor]]

Decorator to be used for every stochastic random variable defined in all statistical models. E.g.:

@bm.random_variable
def foo():
  return Normal(0., 1.)

def foo():
  return Normal(0., 1.)
foo = bm.random_variable(foo)

beanmachine.ppl.model.statistical_model.functional(f: Callable[[beanmachine.ppl.model.statistical_model.P], torch.Tensor]) → Callable[[beanmachine.ppl.model.statistical_model.P], Union[beanmachine.ppl.model.rv_identifier.RVIdentifier, torch.Tensor]]

Decorator to be used for every query defined in statistical model, which are functions of bm.random_variable

@bm.random_variable
def foo():
  return Normal(0., 1.)

@bm.functional():
def bar():
  return foo() * 2.0

beanmachine.ppl.model.statistical_model.param(init_fn)

Decorator to be used for params (variable to be optimized with VI).:

@bm.param
def mu():
  return torch.zeros(2)

@bm.random_variable
def foo():
  return Normal(mu(), 1.)

beanmachine.ppl.model.statistical_model.random_variable(f: Callable[[beanmachine.ppl.model.statistical_model.P], torch.distributions.distribution.Distribution]) → Callable[[beanmachine.ppl.model.statistical_model.P], Union[beanmachine.ppl.model.rv_identifier.RVIdentifier, torch.Tensor]]

Decorator to be used for every stochastic random variable defined in all statistical models. E.g.:

@bm.random_variable
def foo():
  return Normal(0., 1.)

def foo():
  return Normal(0., 1.)
foo = bm.random_variable(foo)