原文：https://github.com/codekansas/keras-language-modeling/blob/master/attention_lstm.py
from __future__ import absolute_import
from keras import backend as K
from keras.engine import InputSpec
from keras.layers import LSTM, activations, Wrapper
class AttentionLSTM(LSTM):
def __init__(self, output_dim, attention_vec, attn_activation='tanh', single_attention_param=False, **kwargs):
self.attention_vec = attention_vec
self.attn_activation = activations.get(attn_activation)
self.single_attention_param = single_attention_param
super(AttentionLSTM, self).__init__(output_dim, **kwargs)
def build(self, input_shape):
super(AttentionLSTM, self).build(input_shape)
if hasattr(self.attention_vec, '_keras_shape'):
attention_dim = self.attention_vec._keras_shape[1]
else:
raise Exception('Layer could not be build: No information about expected input shape.')
self.U_a = self.inner_init((self.output_dim, self.output_dim),
name='{}_U_a'.format(self.name))
self.b_a = K.zeros((self.output_dim,), name='{}_b_a'.format(self.name))
self.U_m = self.inner_init((attention_dim, self.output_dim),
name='{}_U_m'.format(self.name))
self.b_m = K.zeros((self.output_dim,), name='{}_b_m'.format(self.name))
if self.single_attention_param:
self.U_s = self.inner_init((self.output_dim, 1),
name='{}_U_s'.format(self.name))
self.b_s = K.zeros((1,), name='{}_b_s'.format(self.name))
else:
self.U_s = self.inner_init((self.output_dim, self.output_dim),
name='{}_U_s'.format(self.name))
self.b_s = K.zeros((self.output_dim,), name='{}_b_s'.format(self.name))
self.trainable_weights += [self.U_a, self.U_m, self.U_s, self.b_a, self.b_m, self.b_s]
if self.initial_weights is not None:
self.set_weights(self.initial_weights)
del self.initial_weights
def step(self, x, states):
h, [h, c] = super(AttentionLSTM, self).step(x, states)
attention = states[4]
m = self.attn_activation(K.dot(h, self.U_a) * attention + self.b_a)
# Intuitively it makes more sense to use a sigmoid (was getting some NaN problems
# which I think might have been caused by the exponential function -> gradients blow up)
s = K.sigmoid(K.dot(m, self.U_s) + self.b_s)
if self.single_attention_param:
h = h * K.repeat_elements(s, self.output_dim, axis=1)
else:
h = h * s
return h, [h, c]
def get_constants(self, x):
constants = super(AttentionLSTM, self).get_constants(x)
constants.append(K.dot(self.attention_vec, self.U_m) + self.b_m)
return constants
class AttentionLSTMWrapper(Wrapper):
def __init__(self, layer, attention_vec, attn_activation='tanh', single_attention_param=False, **kwargs):
assert isinstance(layer, LSTM)
self.supports_masking = True
self.attention_vec = attention_vec
self.attn_activation = activations.get(attn_activation)
self.single_attention_param = single_attention_param
super(AttentionLSTMWrapper, self).__init__(layer, **kwargs)
def build(self, input_shape):
assert len(input_shape) >= 3
self.input_spec = [InputSpec(shape=input_shape)]
if not self.layer.built:
self.layer.build(input_shape)
self.layer.built = True
super(AttentionLSTMWrapper, self).build()
if hasattr(self.attention_vec, '_keras_shape'):
attention_dim = self.attention_vec._keras_shape[1]
else:
raise Exception('Layer could not be build: No information about expected input shape.')
self.U_a = self.layer.inner_init((self.layer.output_dim, self.layer.output_dim), name='{}_U_a'.format(self.name))
self.b_a = K.zeros((self.layer.output_dim,), name='{}_b_a'.format(self.name))
self.U_m = self.layer.inner_init((attention_dim, self.layer.output_dim), name='{}_U_m'.format(self.name))
self.b_m = K.zeros((self.layer.output_dim,), name='{}_b_m'.format(self.name))
if self.single_attention_param:
self.U_s = self.layer.inner_init((self.layer.output_dim, 1), name='{}_U_s'.format(self.name))
self.b_s = K.zeros((1,), name='{}_b_s'.format(self.name))
else:
self.U_s = self.layer.inner_init((self.layer.output_dim, self.layer.output_dim), name='{}_U_s'.format(self.name))
self.b_s = K.zeros((self.layer.output_dim,), name='{}_b_s'.format(self.name))
self.trainable_weights = [self.U_a, self.U_m, self.U_s, self.b_a, self.b_m, self.b_s]
def get_output_shape_for(self, input_shape):
return self.layer.get_output_shape_for(input_shape)
def step(self, x, states):
h, [h, c] = self.layer.step(x, states)
attention = states[4]
m = self.attn_activation(K.dot(h, self.U_a) * attention + self.b_a)
s = K.sigmoid(K.dot(m, self.U_s) + self.b_s)
if self.single_attention_param:
h = h * K.repeat_elements(s, self.layer.output_dim, axis=1)
else:
h = h * s
return h, [h, c]
def get_constants(self, x):
constants = self.layer.get_constants(x)
constants.append(K.dot(self.attention_vec, self.U_m) + self.b_m)
return constants
def call(self, x, mask=None):
# input shape: (nb_samples, time (padded with zeros), input_dim)
# note that the .build() method of subclasses MUST define
# self.input_spec with a complete input shape.
input_shape = self.input_spec[0].shape
if K._BACKEND == 'tensorflow':
if not input_shape[1]:
raise Exception('When using TensorFlow, you should define '
'explicitly the number of timesteps of '
'your sequences.\n'
'If your first layer is an Embedding, '
'make sure to pass it an "input_length" '
'argument. Otherwise, make sure '
'the first layer has '
'an "input_shape" or "batch_input_shape" '
'argument, including the time axis. '
'Found input shape at layer ' + self.name +
': ' + str(input_shape))
if self.layer.stateful:
initial_states = self.layer.states
else:
initial_states = self.layer.get_initial_states(x)
constants = self.get_constants(x)
preprocessed_input = self.layer.preprocess_input(x)
last_output, outputs, states = K.rnn(self.step, preprocessed_input,
initial_states,
go_backwards=self.layer.go_backwards,
mask=mask,
constants=constants,
unroll=self.layer.unroll,
input_length=input_shape[1])
if self.layer.stateful:
self.updates = []
for i in range(len(states)):
self.updates.append((self.layer.states[i], states[i]))
if self.layer.return_sequences:
return outputs
else:
return last_output