def build(self):
assert self.config['question_len'] == self.config['answer_len']
question = self.question
answer = self.get_answer()
# add embedding layers
weights = self.model_params.get('initial_embed_weights', None)
weights = weights if weights is None else [weights]
embedding = Embedding(input_dim=self.config['n_words'],
output_dim=self.model_params.get(
'n_embed_dims', 100),
weights=weights)
question_embedding = embedding(question)
answer_embedding = embedding(answer)
# turn off layer updating
# embedding.params = []
# embedding.updates = []
# dense
dense = TimeDistributed(
Dense(
self.model_params.get('n_hidden', 200),
# activity_regularizer=regularizers.activity_l1(1e-4),
# W_regularizer=regularizers.l1(1e-4),
activation='tanh'))
question_dense = dense(question_embedding)
answer_dense = dense(answer_embedding)
# cnn
cnns = [
Convolution1D(
filter_length=filter_length,
nb_filter=self.model_params.get('nb_filters', 1000),
activation=self.model_params.get('conv_activation', 'relu'),
# W_regularizer=regularizers.l1(1e-4),
# activity_regularizer=regularizers.activity_l1(1e-4),
border_mode='same') for filter_length in [2, 3, 5, 7]
]
question_cnn = merge([cnn(question_dense) for cnn in cnns],
mode='concat')
answer_cnn = merge([cnn(answer_dense) for cnn in cnns], mode='concat')
# maxpooling
maxpool = Lambda(lambda x: K.max(x, axis=-1, keepdims=False),
output_shape=lambda x: (x[0], x[2]))
avepool = Lambda(lambda x: K.mean(x, axis=-1, keepdims=False),
output_shape=lambda x: (x[0], x[2]))
maxpool.__setattr__('supports_masking', True)
avepool.__setattr__('supports_masking', True)
question_pool = maxpool(question_cnn)
answer_pool = maxpool(answer_cnn)
return question_pool, answer_pool
def build(self):
question = self.question
answer = self.get_answer()
# add embedding layers
weights = self.model_params.get('initial_embed_weights', None)
weights = weights if weights is None else [weights]
embedding = Embedding(
input_dim=self.config['n_words'],
output_dim=self.model_params.get('n_embed_dims', 100),
# W_constraint=constraints.nonneg(),
weights=weights,
mask_zero=True)
question_embedding = embedding(question)
answer_embedding = embedding(answer)
# maxpooling
maxpool = Lambda(lambda x: K.max(x, axis=1, keepdims=False),
output_shape=lambda x: (x[0], x[2]))
maxpool.__setattr__('supports_masking', True)
question_pool = maxpool(question_embedding)
answer_pool = maxpool(answer_embedding)
return question_pool, answer_pool
def build(self):
question = self.question
answer = self.get_answer()
# add embedding layers
weights = self.model_params.get('initial_embed_weights', None)
weights = weights if weights is None else [weights]
embedding = Embedding(
input_dim=self.config['n_words'],
output_dim=self.model_params.get('n_embed_dims', 256),
# weights=weights,
mask_zero=True)
question_embedding = embedding(question)
answer_embedding = embedding(answer)
# turn off layer updating
# embedding.params = []
# embedding.updates = []
# question rnn part
f_rnn = LSTM(self.model_params.get('n_lstm_dims', 141),
return_sequences=True,
dropout_U=0.2,
consume_less='mem')
b_rnn = LSTM(self.model_params.get('n_lstm_dims', 141),
return_sequences=True,
dropout_U=0.2,
consume_less='mem',
go_backwards=True)
question_f_rnn = f_rnn(question_embedding)
question_b_rnn = b_rnn(question_embedding)
# maxpooling
maxpool = Lambda(lambda x: K.max(x, axis=1, keepdims=False),
output_shape=lambda x: (x[0], x[2]))
avepool = Lambda(lambda x: K.mean(x, axis=1, keepdims=False),
output_shape=lambda x: (x[0], x[2]))
# otherwise, it will raise a exception like:
# Layer lambda_1 does not
# support masking, but was passed an input_mask: Elemwise{neq,no_inplace}.0
maxpool.__setattr__('supports_masking', True)
avepool.__setattr__('supports_masking', True)
question_pool = merge(
[maxpool(question_f_rnn),
maxpool(question_b_rnn)],
mode='concat',
concat_axis=-1)
# answer rnn part
f_rnn = AttentionLSTM(self.model_params.get('n_lstm_dims', 141),
question_pool,
return_sequences=True,
consume_less='mem',
single_attention_param=True)
b_rnn = AttentionLSTM(self.model_params.get('n_lstm_dims', 141),
question_pool,
return_sequences=True,
consume_less='mem',
go_backwards=True,
single_attention_param=True)
answer_f_rnn = f_rnn(answer_embedding)
answer_b_rnn = b_rnn(answer_embedding)
answer_pool = merge([maxpool(answer_f_rnn),
maxpool(answer_b_rnn)],
mode='concat',
concat_axis=-1)
return question_pool, answer_pool
