netket.utils.struct.field

netket.utils.struct.field(pytree_node=True, pytree_ignore=False, serialize=None, serialize_name=None, cache=False, sharded=False, **kwargs)

Mark a field of a dataclass or PyTree to be:

Parameters:

• pytree_node (bool) – a leaf node in the pytree representation of this dataclass. If False this must be hashable.

• pytree_ignore (bool) – If True this field is ignored by the pytree metadata and will be excluded from the pytree flattening/unflattening process. This means the field will not appear in the flattened representation when calling jax.tree_util.tree_flatten() or similar pytree operations. This is useful for caches, temporary data, or other fields that should not be passed forward during pytree transformations. When True, pytree_node must be False.

• serialize (bool | None) – Whether this field is saved when the object is serialized (e.g. in a checkpoint). Defaults to pytree_node, so normal fields are saved and static ones are not. Set serialize=False for values recomputed at every step, or serialize=True for a static field that must survive a restart.

• serialize_name (str | None) – If specified, it’s the name under which this attribute is serialized. This can be used to change the runtime attribute name, but maintain some other name in the serialisation format.

• cache (bool) – If True this node is a cache and will be reset every time fields are modified.

• sharded (bool | ShardedFieldSpec) – a boolan or specification object specifying whether this entry is sharded. Defaults to False. If True, a JAX-compatible sharding along axis 0 is assumed.