Settings

Example usage:

# Simple getting and setting.
print(pyro.settings.get())  # print all settings
print(pyro.settings.get("cholesky_relative_jitter"))  # print one
pyro.settings.set(cholesky_relative_jitter=0.5)  # set one
pyro.settings.set(**my_settings)  # set many

# Use as a contextmanager.
with pyro.settings.context(cholesky_relative_jitter=0.5):
    my_function()

# Use as a decorator.
fn = pyro.settings.context(cholesky_relative_jitter=0.5)(my_function)
fn()

# Register a new setting.
pyro.settings.register(
    "binomial_approx_sample_thresh",  # alias
    "pyro.distributions.torch",       # module
    "Binomial.approx_sample_thresh",  # deep name
)

# Register a new setting on a user-provided validator.
@pyro.settings.register(
    "binomial_approx_sample_thresh",  # alias
    "pyro.distributions.torch",       # module
    "Binomial.approx_sample_thresh",  # deep name
)
def validate_thresh(thresh):  # called each time setting is set
    assert isinstance(thresh, float)
    assert thresh > 0

Default Settings

• binomial_approx_log_prob_tol = 0.0

• binomial_approx_sample_thresh = inf

• cholesky_relative_jitter = 4.0

• module_local_params = False

• validate_distributions_pyro = True

• validate_distributions_torch = True

• validate_infer = True

• validate_poutine = True

Settings Interface

get(alias: Optional[str] = None) → Any

Gets one or all global settings.

Parameters

alias (str) – The name of a registered setting.

Returns

The currently set value.

set(**kwargs) → None

Sets one or more settings.

Parameters

**kwargs – alias=value pairs.

context(**kwargs) → Iterator[None]

Context manager to temporarily override one or more settings. This also works as a decorator.

Parameters

**kwargs – alias=value pairs.

register(alias: str, modulename: str, deepname: str, validator: Optional[Callable] = None) → Callable

Register a global settings.

This should be declared in the module where the setting is defined.

This can be used either as a declaration:

settings.register("my_setting", __name__, "MY_SETTING")

or as a decorator on a user-defined validator function:

@settings.register("my_setting", __name__, "MY_SETTING")
def _validate_my_setting(value):
    assert isinstance(value, float)
    assert 0 < value

Parameters

• alias (str) – A valid python identifier serving as a settings alias. Lower snake case preferred, e.g. my_setting.

• modulename (str) – The module name where the setting is declared, typically __name__.

• deepname (str) – A .-separated string of names. E.g. for a module constant, use MY_CONSTANT. For a class attributue, use MyClass.my_attribute.

• validator (callable) – Optional validator that inputs a value, possibly raises validation errors, and returns None.