72 lines
2.0 KiB
Python
72 lines
2.0 KiB
Python
from typing import Callable, Optional, Tuple, cast
|
|
|
|
from ..config import registry
|
|
from ..initializers import glorot_uniform_init, zero_init
|
|
from ..model import Model
|
|
from ..types import Floats1d, Floats2d
|
|
from ..util import get_width, partial
|
|
from .chain import chain
|
|
from .dropout import Dropout
|
|
from .layernorm import LayerNorm
|
|
|
|
InT = Floats2d
|
|
OutT = Floats2d
|
|
|
|
|
|
@registry.layers("Relu.v1")
|
|
def Relu(
|
|
nO: Optional[int] = None,
|
|
nI: Optional[int] = None,
|
|
*,
|
|
init_W: Optional[Callable] = None,
|
|
init_b: Optional[Callable] = None,
|
|
dropout: Optional[float] = None,
|
|
normalize: bool = False,
|
|
) -> Model[InT, OutT]:
|
|
if init_W is None:
|
|
init_W = glorot_uniform_init
|
|
if init_b is None:
|
|
init_b = zero_init
|
|
model: Model[InT, OutT] = Model(
|
|
"relu",
|
|
forward,
|
|
init=partial(init, init_W, init_b),
|
|
dims={"nO": nO, "nI": nI},
|
|
params={"W": None, "b": None},
|
|
)
|
|
if normalize:
|
|
model = chain(model, LayerNorm(nI=nO))
|
|
if dropout is not None:
|
|
model = chain(model, cast(Model[Floats2d, Floats2d], Dropout(dropout)))
|
|
return model
|
|
|
|
|
|
def forward(model: Model[InT, OutT], X: InT, is_train: bool) -> Tuple[OutT, Callable]:
|
|
W = cast(Floats2d, model.get_param("W"))
|
|
b = cast(Floats1d, model.get_param("b"))
|
|
Y = model.ops.affine(X, W, b)
|
|
Y = model.ops.relu(Y)
|
|
|
|
def backprop(dY: OutT) -> InT:
|
|
dY = model.ops.backprop_relu(dY, Y)
|
|
model.inc_grad("b", dY.sum(axis=0))
|
|
model.inc_grad("W", model.ops.gemm(dY, X, trans1=True))
|
|
return model.ops.gemm(dY, W)
|
|
|
|
return Y, backprop
|
|
|
|
|
|
def init(
|
|
init_W: Callable,
|
|
init_b: Callable,
|
|
model: Model[InT, OutT],
|
|
X: Optional[InT] = None,
|
|
Y: Optional[OutT] = None,
|
|
) -> None:
|
|
if X is not None:
|
|
model.set_dim("nI", get_width(X))
|
|
if Y is not None:
|
|
model.set_dim("nO", get_width(Y))
|
|
model.set_param("W", init_W(model.ops, (model.get_dim("nO"), model.get_dim("nI"))))
|
|
model.set_param("b", init_b(model.ops, (model.get_dim("nO"),)))
|