bijx.Const

class bijx.Const

Bases: Variable

Mark a variable as constant during training.

This variable type explicitly indicates that a parameter should remain fixed during optimization. Useful for freezing parts of models or storing hyperparameters that should not be updated.

Can be used with FrozenFilter to selectively freeze parameters during gradient-based optimization.

Example

>>> bijection = SomeBijection()
>>> bijection.scale = Const(1.0)  # Fix scale parameter
>>> # standard optimizers will not apply gradients to scale

__init__(value, *, hijax=None, ref=None, mutable=True, eager_sharding=None, **metadata)

Parameters:

• value (A | VariableMetadata[A])

• hijax (bool | None)

• ref (bool | None)

• mutable (bool)

• eager_sharding (bool | None)

• metadata (Any)

Methods

add_axis(axis_index, axis_name)
copy_from(other)
create_value(value)
del_metadata(name)	Delete a metadata entry for the Variable.
from_metadata(value, attributes)
get_metadata([name, default, exclude_required])	Get metadata for the Variable.
get_raw_value()
get_value(*[, index])
has_metadata(name)	Check if the Variable has a metadata entry for the given name.
remove_axis(axis_index, axis_name)
replace([value, _copy_ref])
set_metadata(*args, **kwargs)	Set metadata for the Variable.
set_raw_value(value, *[, _unsafe_bypass_check])
set_value(value, *[, index])
to_state([value, _copy_ref])
update_from_state(variable_state)

Attributes

hijax
mutable
raw_value
ref
required_metadata
shape
sharding_names
type	The type of the variable.
value
var_type