You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
65 lines
2.0 KiB
65 lines
2.0 KiB
from typing import Callable, List, Optional, Tuple, TypeVar
|
|
|
|
from ..config import registry
|
|
from ..model import Model
|
|
from ..types import Floats1d, Floats2d, Floats3d, Floats4d, FloatsXd, Padded, Ragged
|
|
|
|
# fmt: off
|
|
InT = TypeVar(
|
|
"InT", List[Floats1d], List[Floats2d], List[Floats3d], List[Floats4d],
|
|
Ragged, Padded, FloatsXd, Floats1d, Floats2d, Floats3d, Floats4d)
|
|
# fmt: on
|
|
|
|
|
|
@registry.layers("residual.v1")
|
|
def residual(layer: Model[InT, InT]) -> Model[InT, InT]:
|
|
return Model(
|
|
f"residual({layer.name})",
|
|
forward,
|
|
init=init,
|
|
layers=[layer],
|
|
dims={
|
|
"nO": layer.get_dim("nO") if layer.has_dim("nO") else None,
|
|
"nI": layer.get_dim("nI") if layer.has_dim("nI") else None,
|
|
},
|
|
)
|
|
|
|
|
|
def forward(model: Model[InT, InT], X: InT, is_train: bool) -> Tuple[InT, Callable]:
|
|
def backprop(d_output: InT) -> InT:
|
|
dX = backprop_layer(d_output)
|
|
if isinstance(d_output, list):
|
|
return [d_output[i] + dX[i] for i in range(len(d_output))]
|
|
elif isinstance(d_output, Ragged):
|
|
return Ragged(d_output.data + dX.data, dX.lengths)
|
|
elif isinstance(X, Padded):
|
|
dX.data += d_output.data
|
|
return dX
|
|
else:
|
|
return d_output + dX
|
|
|
|
Y, backprop_layer = model.layers[0](X, is_train)
|
|
if isinstance(X, list):
|
|
return [X[i] + Y[i] for i in range(len(X))], backprop
|
|
elif isinstance(X, Ragged):
|
|
return Ragged(X.data + Y.data, X.lengths), backprop
|
|
elif isinstance(X, Padded):
|
|
Y.data += X.data
|
|
return Y, backprop
|
|
else:
|
|
return X + Y, backprop
|
|
|
|
|
|
def init(
|
|
model: Model[InT, InT], X: Optional[InT] = None, Y: Optional[InT] = None
|
|
) -> None:
|
|
first_layer = model.layers[0]
|
|
if first_layer.has_dim("nO") is None:
|
|
first_layer.initialize(X=X, Y=Y)
|
|
else:
|
|
first_layer.initialize(X=X)
|
|
if first_layer.has_dim("nO"):
|
|
model.set_dim("nO", first_layer.get_dim("nO"))
|
|
if first_layer.has_dim("nI"):
|
|
model.set_dim("nI", first_layer.get_dim("nI"))
|