def build(self, input_shape):
super(MaximumEntropyMarkovModel, self).build(input_shape)
output_dim = input_shape[-1]
if self.hidden_dim is None:
self._trans = self.add_weight(name='trans',
shape=(output_dim, output_dim),
initializer='glorot_uniform',
trainable=True)
if self.lr_multiplier != 1:
K.set_value(self._trans,
K.eval(self._trans) / self.lr_multiplier)
else:
self._l_trans = self.add_weight(name='l_trans',
shape=(output_dim,
self.hidden_dim),
initializer='glorot_uniform',
trainable=True)
self._r_trans = self.add_weight(name='r_trans',
shape=(output_dim,
self.hidden_dim),
initializer='glorot_uniform',
trainable=True)
if self.lr_multiplier != 1:
K.set_value(self._l_trans,
K.eval(self._l_trans) / self.lr_multiplier)
K.set_value(self._r_trans,
K.eval(self._r_trans) / self.lr_multiplier)
def build(self, input_shape):
super(ConditionalRandomField, self).build(input_shape)
output_dim = input_shape[-1]
self._trans = self.add_weight(name='trans',
shape=(output_dim, output_dim),
initializer='glorot_uniform',
trainable=True)
if self.lr_multiplier != 1:
K.set_value(self._trans, K.eval(self._trans) / self.lr_multiplier)
def apply_ema_weights(self, bias_correction=True):
"""备份原模型权重，然后将平均权重应用到模型上去。
"""
self.old_weights = K.batch_get_value(self.model_weights)
ema_weights = K.batch_get_value(self.ema_weights)
if bias_correction:
iterations = K.eval(self.iterations)
scale = 1.0 - np.power(self.ema_momentum, iterations)
ema_weights = [weight / scale for weight in ema_weights]
K.batch_set_value(zip(self.model_weights, ema_weights))