def _get_sog_activation(self, x, s, configs, scope, name):
assert isinstance(configs, (list, tuple)) and len(configs) > 0
assert isinstance(self._gate_signals, list) and len(
self._gate_signals) == 3
net_u = self._hyper_neuron_16(x, s, self._gate_signals, scope)
u = linker.softmax_over_groups(net_u, configs, name)
return u
def _get_sog_activation(self, x, s, configs, scope, name):
assert isinstance(configs, (list, tuple)) and len(configs) > 0
net_u = self.dense_rn(x, s, scope)
if th.sog_version == 0:
u = linker.softmax_over_groups(net_u, configs, name)
else:
u = self._sog_v1(net_u)
return u
def _get_coupled_gates(self, x, configs, reverse):
assert isinstance(configs, (list, tuple))
# T for transform, C for carry
net_T = self.neurons(x, scope='net_u')
if len(configs) == 0: T = tf.sigmoid(net_T, name='transform_gate')
else: T = linker.softmax_over_groups(net_T, configs, 'transform_gate')
C = tf.subtract(1., T, name='carry_gate')
if len(configs) > 0 and reverse: T, C = C, T
if hub.export_gates: self.tensors_to_export['carry_gate'] = C
return T, C
def _get_sog_activation(self, x, s, configs, scope, name):
assert isinstance(configs, (list, tuple)) and len(configs) > 0
net_u = self.neurons(x, s, scope=scope)
u = linker.softmax_over_groups(net_u, configs, name)
return u