def lstm_cell(units: int,
peephole: bool = True,
kernel_activation: str = "tanh",
recurrent_activation: str = "hard_sigmoid",
kernel_initializer: str = "glorot_uniform",
recurrent_initializer: str = "orthogonal",
bias_initializer: str = "zeros",
use_bias: bool = True,
**kwargs) -> RNNCellBuilder:
cell = RNNCellBuilder(units, ["h", "C"], kernel_activation,
recurrent_activation, kernel_initializer,
recurrent_initializer, bias_initializer, use_bias,
**kwargs)
if peephole:
return cell \
.add_recurrent("input", ["X", "h", "C"]) \
.add_recurrent("forget", ["X", "h", "C"]) \
.add_recurrent("output", ["X", "h", "C"]) \
.add_kernel("cell", ["X", "h"]) \
.add_var("C_next", ["forget", "C", "input", "cell"],
lambda x: cell.recurrent_activation(x[0] * x[1] + x[2] * x[3])) \
.add_var("h_next", ["C_next", "output"], lambda x: cell.kernel_activation(x[0]) * x[1])
return cell \
.add_recurrent("input", ["X", "h"]) \
.add_recurrent("forget", ["X", "h"]) \
.add_recurrent("output", ["X", "h"]) \
.add_kernel("cell", ["X", "h"]) \
.add_var("C_next", ["forget", "C", "input", "cell"],
lambda x: cell.recurrent_activation(x[0] * x[1] + x[2] * x[3])) \
.add_var("h_next", ["C_next", "output"], lambda x: cell.kernel_activation(x[0]) * x[1])
def gru_cell(units: int,
kernel_activation: str = "tanh",
recurrent_activation: str = "hard_sigmoid",
kernel_initializer: str = "glorot_uniform",
recurrent_initializer: str = "orthogonal",
bias_initializer: str = "zeros",
use_bias: bool = True,
**kwargs) -> RNNCellBuilder:
cell = RNNCellBuilder(units, ["h"], kernel_activation,
recurrent_activation, kernel_initializer,
recurrent_initializer, bias_initializer, use_bias,
**kwargs)
return cell \
.add_recurrent("update", ["X", "h"]) \
.add_recurrent("reset", ["X", "h"]) \
.add_kernel("output", ["X", "reset"]) \
.add_var("h_next", ["update", "h", "output"], lambda x: (1 - x[0]) * x[1] + x[0] * x[2])