def _neurons_forward_with_memory(self,
x,
s,
name,
f,
fc_mem,
output_dim=None,
use_bias=True,
truncate=False,
w_reg=None):
# If fully connect memory
if fc_mem:
return self._neurons_forward(tf.concat([x, s], axis=1),
name,
f,
output_dim,
use_bias,
truncate,
w_reg=w_reg)
# Otherwise
dim = self._state_size if output_dim is None else output_dim
return linker.neurons(dim,
x,
activation=f,
memory=s,
fc_memory=fc_mem,
use_bias=use_bias,
scope=name,
truncate=truncate,
weight_regularizer=w_reg)
def _neurons_forward(
self, x, name, f, output_dim=None, use_bias=True, truncate=False,
w_reg=None):
assert name is not None and callable(f)
dim = self._state_size if output_dim is None else output_dim
return linker.neurons(
dim, x, activation=f, use_bias=use_bias, truncate=truncate, scope=name,
weight_regularizer=w_reg)
def neurons(self,
x,
s=None,
num=None,
fc_memory=True,
is_gate=False,
activation=None,
scope=None,
truncate=False,
num_or_size_splits=None,
weight_initializer=None,
use_bias=None,
bias_initializer=None,
weight_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
**kwargs):
if num is None:
if isinstance(num_or_size_splits, int):
num = num_or_size_splits * self._state_size
elif isinstance(num_or_size_splits, (list, tuple)):
num = sum(num_or_size_splits)
else:
num = self._state_size
if activation is None and is_gate:
activation = tf.sigmoid
if weight_initializer is None:
weight_initializer = getattr(self, '_weight_initializer', None)
if use_bias is None:
use_bias = getattr(self, '_use_bias', None)
if bias_initializer is None:
bias_initializer = getattr(self, '_bias_initializer')
if weight_regularizer is None:
weight_regularizer = getattr(self, '_weight_regularizer', None)
if bias_regularizer is None:
bias_regularizer = getattr(self, '_bias_regularizer', None)
if activity_regularizer is None:
activity_regularizer = getattr(self, '_activity_regularizer', None)
_kwargs = getattr(self, '_kwargs', {})
return linker.neurons(num=num,
external_input=x,
activation=activation,
memory=s,
fc_memory=fc_memory,
scope=scope,
use_bias=use_bias,
truncate=truncate,
num_or_size_splits=num_or_size_splits,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
weight_regularizer=weight_regularizer,
bias_regularizer=bias_regularizer,
activity_regularizer=activity_regularizer,
**_kwargs,
**kwargs)
def neurons(self,
x,
num=None,
is_gate=False,
activation=None,
scope=None,
truncate=False,
num_or_size_splits=None,
weight_initializer=None,
use_bias=None,
bias_initializer=None,
weight_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
prune_frac=0,
**kwargs):
if num is None:
if isinstance(num_or_size_splits, int):
assert self._output_dim is not None
num = num_or_size_splits * self._output_dim
elif isinstance(num_or_size_splits, (list, tuple)):
num = sum(num_or_size_splits)
else:
assert self._output_dim is not None
num = self._output_dim
if activation is None and is_gate:
activation = tf.sigmoid
if weight_initializer is None:
weight_initializer = getattr(self, '_weight_initializer', None)
if use_bias is None:
use_bias = getattr(self, '_use_bias', None)
if bias_initializer is None:
bias_initializer = getattr(self, '_bias_initializer')
if weight_regularizer is None:
weight_regularizer = getattr(self, '_weight_regularizer', None)
if bias_regularizer is None:
bias_regularizer = getattr(self, '_bias_regularizer', None)
if activity_regularizer is None:
activity_regularizer = getattr(self, '_activity_regularizer', None)
return linker.neurons(num=num,
external_input=x,
activation=activation,
scope=scope,
use_bias=use_bias,
truncate=truncate,
num_or_size_splits=num_or_size_splits,
weight_initializer=weight_initializer,
bias_initializer=bias_initializer,
weight_regularizer=weight_regularizer,
bias_regularizer=bias_regularizer,
activity_regularizer=activity_regularizer,
prune_frac=prune_frac,
**kwargs)
