def tag_recurrent_dropout(self,
variables,
recurrent_dropout,
rng=None,
**hyperparameters):
from blocks.roles import OUTPUT, has_roles
ancestors = graph.deep_ancestors(variables)
for lstm in self.rnn.transitions:
variables = [
var for var in ancestors
if (has_roles(var, [OUTPUT]) and lstm in var.tag.annotations
and var.name.endswith("states"))
]
# get one dropout mask for all time steps. use the very
# first state to get the hidden state shape, else we get
# graph cycles.
initial_state = util.the(
[var for var in variables if "initial_state" in var.name])
mask = util.get_dropout_mask(initial_state.shape,
recurrent_dropout,
rng=rng)
subsequent_states = [
var for var in variables if "initial_state" not in var.name
]
graph.add_transform(subsequent_states,
graph.DropoutTransform("recurrent_dropout",
mask=mask),
reason="regularization")
def __call__(self, x, rng=None, **hyperparameters):
p = hyperparameters[self.key]
if isinstance(p, numbers.Number) and p <= 0:
return x
mask = self.mask or util.get_dropout_mask(
x.shape, p, rng=self.rng or rng)
return x * mask
