"""Sparse synapse layer registry and construction helpers."""
from __future__ import annotations
import torch.nn as nn
from dendritic_modeling.config.conversion import normalize_sparsity_type
from dendritic_modeling.networks.architectures.excitation_inhibition.synapse.credit_deepst import (
CreditGatedDeepstLinear,
)
from dendritic_modeling.networks.architectures.excitation_inhibition.synapse.deepst import (
DeepstLinear,
)
from dendritic_modeling.networks.architectures.excitation_inhibition.synapse.dense_to_sparse import (
DenseToSparseLinear,
)
from dendritic_modeling.networks.architectures.excitation_inhibition.synapse.indexed import (
IndexedDynamicTopKLinear,
IndexedRewireLinear,
IndexedSparseLinear,
)
from dendritic_modeling.networks.architectures.excitation_inhibition.synapse.indexed_release import (
IndexedReleaseLinear,
)
from dendritic_modeling.networks.architectures.excitation_inhibition.synapse.stochastic import (
StochasticTopKLinear,
)
from dendritic_modeling.networks.architectures.excitation_inhibition.synapse.topk import (
TopKLinear,
)
from dendritic_modeling.networks.architectures.excitation_inhibition.synapse.variance import (
VarianceTopKLinear,
)
from dendritic_modeling.plugins import load_plugins
SPARSE_LAYER_REGISTRY: dict[str, type[nn.Module]] = {
"indexed_release": IndexedReleaseLinear,
"standard": TopKLinear,
"stochastic": StochasticTopKLinear,
"variance": VarianceTopKLinear,
"deepst": DeepstLinear,
"credit_deepst": CreditGatedDeepstLinear,
"dense_to_sparse": DenseToSparseLinear,
"annealed_topk": DenseToSparseLinear,
"indexed": IndexedSparseLinear,
"indexed_dynamic": IndexedDynamicTopKLinear,
"indexed_rewire": IndexedRewireLinear,
}
#: The types DendriNet registers itself; any other registered type is a plugin's.
BUILTIN_SPARSE_LAYER_TYPES = frozenset(SPARSE_LAYER_REGISTRY)
def normalize_sparse_layer_type(layer_type: object) -> str:
"""Normalize public sparse-layer aliases to registry keys."""
return normalize_sparsity_type(layer_type)
def get_sparse_layer_class(layer_type: str) -> type[nn.Module]:
"""The class registered under ``layer_type``, loading plugins for an unknown name."""
layer_type = normalize_sparse_layer_type(layer_type)
if layer_type not in SPARSE_LAYER_REGISTRY:
load_plugins()
if layer_type not in SPARSE_LAYER_REGISTRY:
available_types = list(SPARSE_LAYER_REGISTRY.keys())
raise ValueError(
f"Unknown sparse layer type '{layer_type}'. Available types: {available_types}"
)
return SPARSE_LAYER_REGISTRY[layer_type]
def get_sparse_layer(
layer_type: str,
in_features: int,
out_features: int,
K: int,
**kwargs,
) -> nn.Module:
"""Create a sparse synapse layer by registered type."""
layer_type = normalize_sparse_layer_type(layer_type)
layer_class = get_sparse_layer_class(layer_type)
if layer_type in {"deepst", "credit_deepst"}:
return layer_class(in_features=in_features, out_features=out_features, **kwargs)
return layer_class(
in_features=in_features,
out_features=out_features,
K=K,
**kwargs,
)
[docs]
def register_sparse_layer(
name: str, layer_class: type[nn.Module], *, allow_override: bool = False
) -> None:
"""Register a sparse synapse implementation under a new type name.
Replacing an existing type (built-in or registered) requires
``allow_override=True``.
"""
if not str(name).strip():
raise ValueError("sparse layer type name must be non-empty")
key = normalize_sparse_layer_type(str(name).strip())
if key in SPARSE_LAYER_REGISTRY and not allow_override:
raise ValueError(
f"Sparse layer type '{key}' is already registered; pass "
"allow_override=True to replace it"
)
SPARSE_LAYER_REGISTRY[key] = layer_class
def unregister_sparse_layer(name: str) -> None:
"""Remove a registered sparse synapse type if present."""
SPARSE_LAYER_REGISTRY.pop(normalize_sparse_layer_type(str(name).strip()), None)
def get_available_sparse_types() -> list[str]:
"""Return registered sparse synapse type names, including those of plugins."""
load_plugins()
return list(SPARSE_LAYER_REGISTRY.keys())
__all__ = [
"BUILTIN_SPARSE_LAYER_TYPES",
"SPARSE_LAYER_REGISTRY",
"get_available_sparse_types",
"get_sparse_layer",
"get_sparse_layer_class",
"normalize_sparse_layer_type",
"register_sparse_layer",
"unregister_sparse_layer",
]