def build(self):
seq1 = Input(name='seq1', shape=[self.config['seq1_maxlen']])
seq2 = Input(name='seq2', shape=[self.config['seq2_maxlen']])
embedding = Embedding(self.config['vocab_size'], self.config['embed_size'], weights=[self.config['embed']], trainable = self.config['embed_trainable'])
seq1_embed = embedding(seq1)
seq1_embed = Dropout(0.5)(seq1_embed)
seq2_embed = embedding(seq2)
seq2_embed = Dropout(0.5)(seq2_embed)
lstm = Bidirectional(LSTM(self.config['hidden_size'], return_sequences=True, dropout=self.config['dropout_rate']))
seq1_rep = lstm(seq1_embed)
seq2_rep = lstm(seq2_embed)
cross = Match(match_type='concat')([seq1_rep, seq2_rep])
cross_reshape = Reshape((-1, ))(cross)
mm_k = Lambda(lambda x: K.tf.nn.top_k(x, k=100, sorted=True)[0])(cross_reshape)
pool_flat_drop = Dropout(rate=self.config['dropout_rate'])(mm_k)
if self.config['target_mode'] == 'classification':
out = Dense(2, activation='softmax')(pool_flat_drop)
elif self.config['target_mode'] in ['regression', 'ranking']:
out = Dense(1)(pool_flat_drop)
model = Model(inputs=[seq1, seq2], outputs=out)
return model
def build_model(self):
encoding_layer1 = Bidirectional(GRU(300,
return_sequences=True,
dropout=0.2))
encoded_sentence_1 = encoding_layer1(self.Q1_emb) # (?, len, 600)
encoded_sentence_2 = encoding_layer1(self.Q2_emb) # (?, len, 600)
q_conv1 = Conv1D(32, 3, padding='same')(encoded_sentence_1)
show_layer_info('Conv1D', q_conv1)
d_conv1 = Conv1D(32, 3, padding='same')(encoded_sentence_2)
show_layer_info('Conv1D', d_conv1)
cross = Match(match_type='plus')([q_conv1, d_conv1])
show_layer_info('Match-plus', cross)
z = Reshape((self.word_max_len, self.word_max_len, -1))(cross)
show_layer_info('Reshape', z)
for i in range(2):
z = Conv2D(filters=self.kernel_counts_2d[i], kernel_size=self.kernel_sizes_2d[i], padding='same', activation='relu')(z)
show_layer_info('Conv2D', z)
z = MaxPooling2D(pool_size=(self.mpool_sizes_2d[i][0], self.mpool_sizes_2d[i][1]))(z)
show_layer_info('MaxPooling2D', z)
pool1_flat = Flatten()(z)
show_layer_info('Flatten', pool1_flat)
pool1_flat_drop = Dropout(rate=0.2)(pool1_flat)
show_layer_info('Dropout', pool1_flat_drop)
out_ = Dense(1, activation='sigmoid')(pool1_flat_drop)
show_layer_info('Dense', out_)
return out_
def build(self):
query = Input(name='query', shape=(self.config['text1_maxlen'], ))
show_layer_info('Input', query)
doc = Input(name='doc', shape=(self.config['text2_maxlen'], ))
show_layer_info('Input', doc)
embedding = Embedding(self.config['vocab_size'],
self.config['embed_size'],
weights=[self.config['embed']],
trainable=self.embed_trainable)
q_embed = embedding(query)
show_layer_info('Embedding', q_embed)
d_embed = embedding(doc)
show_layer_info('Embedding', d_embed)
q_rep = Bidirectional(
LSTM(self.config['hidden_size'],
return_sequences=True,
dropout=self.config['dropout_rate']))(q_embed)
show_layer_info('Bidirectional-LSTM', q_rep)
d_rep = Bidirectional(
LSTM(self.config['hidden_size'],
return_sequences=True,
dropout=self.config['dropout_rate']))(d_embed)
show_layer_info('Bidirectional-LSTM', d_rep)
cross = Match(match_type='dot')([q_rep, d_rep])
#cross = Dot(axes=[2, 2])([q_embed, d_embed])
show_layer_info('Match-dot', cross)
cross_reshape = Reshape((-1, ))(cross)
show_layer_info('Reshape', cross_reshape)
mm_k = Lambda(lambda x: K.tf.nn.top_k(
x, k=self.config['topk'], sorted=True)[0])(cross_reshape)
show_layer_info('Lambda-topk', mm_k)
pool1_flat_drop = Dropout(rate=self.config['dropout_rate'])(mm_k)
show_layer_info('Dropout', pool1_flat_drop)
if self.config['target_mode'] == 'classification':
out_ = Dense(2, activation='softmax')(pool1_flat_drop)
elif self.config['target_mode'] in ['regression', 'ranking']:
out_ = Dense(1)(pool1_flat_drop)
show_layer_info('Dense', out_)
#model = Model(inputs=[query, doc, dpool_index], outputs=out_)
model = Model(inputs=[query, doc], outputs=out_)
model.summary()
return model
def build(self):
def conv2d_work(input_dim):
seq = Sequential()
assert self.config['num_conv2d_layers'] > 0
for i in range(self.config['num_conv2d_layers']):
seq.add(Conv2D(filters=self.config['2d_kernel_counts'][i], kernel_size=self.config['2d_kernel_sizes'][i], padding='same', activation='relu'))
seq.add(MaxPooling2D(pool_size=(self.config['2d_mpool_sizes'][i][0], self.config['2d_mpool_sizes'][i][1])))
return seq
query = Input(name='query', shape=(self.config['text1_maxlen'],))
show_layer_info('Input', query)
doc = Input(name='doc', shape=(self.config['text2_maxlen'],))
show_layer_info('Input', doc)
embedding = Embedding(self.config['vocab_size'], self.config['embed_size'], weights=[self.config['embed']], trainable = self.embed_trainable)
q_embed = embedding(query)
show_layer_info('Embedding', q_embed)
d_embed = embedding(doc)
show_layer_info('Embedding', d_embed)
q_conv1 = Conv1D(self.config['1d_kernel_count'], self.config['1d_kernel_size'], padding='same') (q_embed)
show_layer_info('Conv1D', q_conv1)
d_conv1 = Conv1D(self.config['1d_kernel_count'], self.config['1d_kernel_size'], padding='same') (d_embed)
show_layer_info('Conv1D', d_conv1)
cross = Match(match_type='plus')([q_conv1, d_conv1])
show_layer_info('Match-plus', cross)
z = Reshape((self.config['text1_maxlen'], self.config['text2_maxlen'], -1))(cross)
show_layer_info('Reshape', z)
for i in range(self.config['num_conv2d_layers']):
z = Conv2D(filters=self.config['2d_kernel_counts'][i], kernel_size=self.config['2d_kernel_sizes'][i], padding='same', activation='relu')(z)
show_layer_info('Conv2D', z)
z = MaxPooling2D(pool_size=(self.config['2d_mpool_sizes'][i][0], self.config['2d_mpool_sizes'][i][1]))(z)
show_layer_info('MaxPooling2D', z)
#dpool = DynamicMaxPooling(self.config['dpool_size'][0], self.config['dpool_size'][1])([conv2d, dpool_index])
pool1_flat = Flatten()(z)
show_layer_info('Flatten', pool1_flat)
pool1_flat_drop = Dropout(rate=self.config['dropout_rate'])(pool1_flat)
show_layer_info('Dropout', pool1_flat_drop)
if self.config['target_mode'] == 'classification':
out_ = Dense(2, activation='softmax')(pool1_flat_drop)
elif self.config['target_mode'] in ['regression', 'ranking']:
out_ = Dense(1)(pool1_flat_drop)
show_layer_info('Dense', out_)
model = Model(inputs=[query, doc], outputs=out_)
return model
def build(self):
def conv2d_work(input_dim):
seq = Sequential()
assert self.config['num_conv2d_layers'] > 0
for i in range(self.config['num_conv2d_layers']):
seq.add(
Conv2D(filters=self.config['2d_kernel_counts'][i],
kernel_size=self.config['2d_kernel_sizes'][i],
padding='same',
activation='relu'))
seq.add(
MaxPooling2D(
pool_size=(self.config['2d_mpool_sizes'][i][0],
self.config['2d_mpool_sizes'][i][1])))
return seq
query = Input(name='query', shape=(self.config['text1_maxlen'], ))
show_layer_info('Input', query)
doc = Input(name='doc', shape=(self.config['text2_maxlen'], ))
show_layer_info('Input', doc)
embedding = Embedding(self.config['vocab_size'],
self.config['embed_size'],
weights=[self.config['embed']],
trainable=self.embed_trainable)
q_embed = embedding(query)
show_layer_info('Embedding', q_embed)
d_embed = embedding(doc)
show_layer_info('Embedding', d_embed)
# ########## compute attention weights for the query words: better then mvlstm alone
if self.config["text1_attention"]:
q_w = Dense(1,
kernel_initializer=self.initializer_gate,
use_bias=False)(
q_embed) # use_bias=False to simple combination
show_layer_info('Dense', q_w)
q_w = Lambda(lambda x: softmax(x, axis=1),
output_shape=(self.config['text1_maxlen'], ),
name="q_w")(q_w)
show_layer_info('Lambda-softmax', q_w)
# ########## add attention weights for Q_words
q_w_layer = Lambda(lambda x: K.repeat_elements(
q_w, rep=self.config['embed_size'], axis=2))(q_w)
show_layer_info('repeat', q_w_layer)
q_embed = Multiply()([q_w_layer, q_embed])
show_layer_info('Dot-qw', q_embed)
# ####################### attention text1
# ########## compute attention weights for the document words:
if self.config['text2_attention']:
d_w = Dense(1,
kernel_initializer=self.initializer_gate,
use_bias=False)(d_embed)
show_layer_info('Dense', d_w)
d_w = Lambda(lambda x: softmax(x, axis=1),
output_shape=(self.config['text2_maxlen'], ),
name="d_w")(d_w)
show_layer_info('Lambda-softmax', d_w)
# ########## add attention weights for D_words
d_w_layer = Lambda(lambda x: K.repeat_elements(
d_w, rep=self.config['embed_size'], axis=2))(d_w)
d_embed = Multiply()([d_w_layer, d_embed])
show_layer_info('Dot-qw', d_embed)
# ####################### attention text2
q_conv1 = Conv1D(self.config['1d_kernel_count'],
self.config['1d_kernel_size'],
padding='same')(q_embed)
show_layer_info('Conv1D', q_conv1)
d_conv1 = Conv1D(self.config['1d_kernel_count'],
self.config['1d_kernel_size'],
padding='same')(d_embed)
show_layer_info('Conv1D', d_conv1)
cross = Match(match_type='plus')([q_conv1, d_conv1])
show_layer_info('Match-plus', cross)
z = Reshape((self.config['text1_maxlen'], self.config['text2_maxlen'],
-1))(cross)
show_layer_info('Reshape', z)
# add the passages attention
if self.config["passage_attention"]:
# ########################## compute the passages attention weights
p_cross = Permute((2, 1, 3))(z)
show_layer_info('p_cross', p_cross)
starts = [
i for i in range(0, self.config['text2_maxlen'],
self.config['context_len'])
]
slice_layer = [
crop(1, start, start + self.config['context_len'])
for start in starts
]
slices = [slice_layer_i(p_cross) for slice_layer_i in slice_layer]
attention_ws = []
for slice in slices:
s_dw = Dense(1, use_bias=False)(slice)
s_dw = Lambda(lambda x: softmax(x, axis=1))(s_dw)
attention_ws.append(s_dw)
d_w = concatenate(attention_ws, 1)
show_layer_info('attW', d_w)
z = Multiply()([d_w, p_cross])
show_layer_info('Multiply', z)
# ########################## passages attention
for i in range(self.config['num_conv2d_layers']):
z = Conv2D(filters=self.config['2d_kernel_counts'][i],
kernel_size=self.config['2d_kernel_sizes'][i],
padding='same',
activation='relu')(z)
show_layer_info('Conv2D', z)
z = MaxPooling2D(
pool_size=(self.config['2d_mpool_sizes'][i][0],
self.config['2d_mpool_sizes'][i][1]))(z)
show_layer_info('MaxPooling2D', z)
#dpool = DynamicMaxPooling(self.config['dpool_size'][0], self.config['dpool_size'][1])([conv2d, dpool_index])
pool1_flat = Flatten()(z)
show_layer_info('Flatten', pool1_flat)
pool1_flat_drop = Dropout(rate=self.config['dropout_rate'])(pool1_flat)
show_layer_info('Dropout', pool1_flat_drop)
if self.config['target_mode'] == 'classification':
out_ = Dense(2, activation='softmax')(pool1_flat_drop)
elif self.config['target_mode'] in ['regression', 'ranking']:
out_ = Dense(1)(pool1_flat_drop)
show_layer_info('Dense', out_)
model = Model(inputs=[query, doc], outputs=out_)
return model
def build(self):
query = Input(name="query", batch_shape=[None, None], dtype='int32')
show_layer_info('Input', query)
doc = Input(name="doc", batch_shape=[None, None], dtype='int32')
show_layer_info('Input', doc)
input_embed = self.config['vocab_size'] if self.config['mask_zero'] else self.config['vocab_size']
embedding = Embedding(input_embed, self.config['embed_size'], weights=[self.config['embed']],
trainable=self.embed_trainable, name="embeddings",
mask_zero=self.config['mask_zero'])
q_embed = embedding(query)
show_layer_info('Embedding', q_embed)
d_embed = embedding(doc)
show_layer_info('Embedding', d_embed)
q_lstm_layer = Bidirectional(LSTM(self.config["number_q_lstm_units"],
dropout=self.config["q_lstm_dropout"],
recurrent_dropout=self.config["q_lstm_dropout"],
return_sequences=True),
name="q_lstm")
d_lstm_layer = Bidirectional(LSTM(self.config["number_d_lstm_units"],
dropout=self.config["d_lstm_dropout"],
recurrent_dropout=self.config["d_lstm_dropout"],
return_sequences=True),
name="d_lstm")
q_mat = q_lstm_layer(q_embed)
show_layer_info('Bibirectional-LSTM', q_mat)
d_mat = d_lstm_layer(d_embed)
show_layer_info('Bibirectional-LSTM', d_mat)
input_mat = Match(normalize=True)([q_mat, d_mat]) # the result is cosine similarity matrix
show_layer_info('Match', input_mat)
#input_mat = BatchNormalization()(input_mat)
#input_mat = Dropout(self.config["dropout_rate"])(input_mat)
input_mat = Reshape((self.config["text1_maxlen"], self.config["text2_maxlen"]))(input_mat)
show_layer_info('Match', input_mat)
merged = Conv1D(self.config['filters'], self.config['kernel_size'],
activation=self.config['conv_activation'], name="conv1", padding='same')(input_mat)
merged = BatchNormalization()(merged)
merged = Dropout(self.config["conv_dropout"])(merged)
show_layer_info('Conv1D', merged)
merged = MaxPooling1D(pool_size=self.config['pool_size'], name="maxPool1")(merged)
show_layer_info('MaxPooling1D', merged)
merged = Conv1D(self.config['filters'], self.config['kernel_size'],
activation=self.config['conv_activation'], name="conv2", padding='same')(input_mat)
show_layer_info('Conv1D', merged)
merged = BatchNormalization()(merged)
merged = Dropout(self.config["conv_dropout"])(merged)
merged = MaxPooling1D(pool_size=self.config['pool_size'], name="maxPool2")(merged)
show_layer_info('MaxPooling1D', merged)
"""
merged = Conv1D(self.config['filters'], self.config['kernel_size'],
activation=self.config['conv_activation'], name="conv3", padding='same')(input_mat)
show_layer_info('Conv1D', merged)
merged = BatchNormalization()(merged)
merged = Dropout(self.config["conv_dropout"])(merged)
merged = MaxPooling1D(pool_size=self.config['pool_size'], name="maxPool3")(merged)
"""
show_layer_info('MaxPooling1D', merged)
merged = Flatten()(merged)
dense = Dense(self.config["hidden_sizes"][0], activation=self.config['hidden_activation'],
name="MLP_combine_0")(merged)
show_layer_info('Dense', dense)
for i in range(self.config["num_layers"] - 1):
dense = BatchNormalization()(dense)
dense = Dropout(self.config["dropout_rate"])(dense)
dense = Dense(self.config["hidden_sizes"][i + 1], activation=self.config['hidden_activation'],
name="MLP_combine_" + str(i + 1))(dense)
show_layer_info('Dense', dense)
dense = BatchNormalization()(dense)
dense = Dropout(self.config["dropout_rate"])(dense)
# out_ = Dense(1, activation=self.config['output_activation'], name="MLP_out")(dense)
if self.config['target_mode'] == 'classification':
out_ = Dense(2, activation=self.config['output_activation'], name="MLP_out")(dense)
elif self.config['target_mode'] in ['regression', 'ranking']:
out_ = Dense(1, activation=self.config['output_activation'], name="MLP_out")(dense)
show_layer_info('Output', out_)
model = Model(inputs=[query, doc], outputs=[out_])
plot_model(model, to_file='../conv_wc_model_plot.png', show_shapes=True, show_layer_names=True)
return model
def build(self):
query = Input(name='query', shape=(self.config['text1_maxlen'], ))
show_layer_info('Input', query)
doc = Input(name='doc', shape=(self.config['text2_maxlen'], ))
show_layer_info('Input', doc)
embedding = Embedding(self.config['vocab_size'],
self.config['embed_size'],
weights=[self.config['embed']],
trainable=self.embed_trainable)
q_embed = embedding(query)
show_layer_info('Embedding_q', q_embed)
# ########## compute attention weights for the query words: better then mvlstm alone
if self.config["text1_attention"]:
q_w = Dense(1,
kernel_initializer=self.initializer_gate,
use_bias=False)(
q_embed) # use_bias=False to simple combination
show_layer_info('Dense', q_w)
q_w = Lambda(lambda x: softmax(x, axis=1),
output_shape=(self.config['text1_maxlen'], ))(q_w)
show_layer_info('Lambda-softmax', q_w)
# ########## add attention weights for Q_words
q_w_layer = Lambda(lambda x: K.repeat_elements(
q_w, rep=self.config['embed_size'], axis=2))(q_w)
show_layer_info('repeat', q_w_layer)
q_embed = Multiply()([q_w_layer, q_embed])
show_layer_info('Dot-qw', q_embed)
# ####################### attention
d_embed = embedding(doc)
show_layer_info('Embedding_d', d_embed)
# ########## compute attention weights for the document words:
if self.config['text2_attention']:
d_w = Dense(1,
kernel_initializer=self.initializer_gate,
use_bias=False)(d_embed)
show_layer_info('Dense', d_w)
d_w = Lambda(lambda x: softmax(x, axis=1),
output_shape=(self.config['text2_maxlen'], ))(d_w)
show_layer_info('Lambda-softmax', d_w)
# ########## add attention weights for D_words
d_w_layer = Lambda(lambda x: K.repeat_elements(
d_w, rep=self.config['embed_size'], axis=2))(d_w)
d_embed = Multiply()([d_w_layer, d_embed])
show_layer_info('Dot-qw', d_embed)
# ####################### attention
q_rep = Bidirectional(
LSTM(self.config['hidden_size'],
return_sequences=True,
dropout=self.config['dropout_rate']))(q_embed)
show_layer_info('Bidirectional-LSTM_q', q_rep)
q_rep = BatchNormalization()(q_rep)
q_rep = Dropout(self.config["dropout_lstm"])(q_rep)
d_rep = Bidirectional(
LSTM(self.config['hidden_size'],
return_sequences=True,
dropout=self.config['dropout_rate']))(d_embed)
show_layer_info('Bidirectional-LSTM_d', d_rep)
d_rep = BatchNormalization()(d_rep)
d_rep = Dropout(self.config["dropout_lstm"])(d_rep)
cross = Match(match_type='dot')([q_rep, d_rep])
show_layer_info('Match-dot', cross)
# ####################### convolutions
cov1 = Conv2D(self.config['filters'],
self.config['kernel_size'],
activation='relu',
name="conv1",
padding='same')(cross)
cov1 = BatchNormalization()(cov1)
cov1 = Dropout(self.config["dropout_rate"])(cov1)
show_layer_info('Conv1', cov1)
cov1 = MaxPooling2D(pool_size=3, name="maxPool")(cov1)
show_layer_info('MaxPooling1D-cov2', cov1)
cov2 = Conv2D(self.config['filters'],
self.config['kernel_size'],
activation='relu',
name="conv2",
padding='same')(cov1)
cov2 = BatchNormalization()(cov2)
cross = Dropout(self.config["dropout_rate"])(cov2)
show_layer_info('Conv2', cov2)
# ###################### convolutions
cross_reshape = Reshape((-1, ))(cross)
show_layer_info('Reshape', cross_reshape)
mm_k = Lambda(lambda x: K.tf.nn.top_k(
x, k=self.config['topk'], sorted=True)[0])(cross_reshape)
show_layer_info('Lambda-topk', mm_k)
pool1_flat_drop = Dropout(rate=self.config['dropout_rate'])(mm_k)
show_layer_info('Dropout', pool1_flat_drop)
if self.config['target_mode'] == 'classification':
out_ = Dense(2, activation='softmax')(pool1_flat_drop)
elif self.config['target_mode'] in ['regression', 'ranking']:
out_ = Dense(1)(pool1_flat_drop)
show_layer_info('Dense', out_)
model = Model(inputs=[query, doc], outputs=out_)
plot_model(model,
to_file='../amvlstm_conv.png',
show_shapes=True,
show_layer_names=True)
return model
def build(self):
query = Input(name='query', shape=(self.config['text1_maxlen'], ))
show_layer_info('Input', query)
doc = Input(name='doc', shape=(self.config['text2_maxlen'], ))
show_layer_info('Input', doc)
embedding = Embedding(self.config['vocab_size'],
self.config['embed_size'],
weights=[self.config['embed']],
trainable=self.embed_trainable)
q_embed = embedding(query)
show_layer_info('Embedding_q', q_embed)
# ########## compute attention weights for the query words: better then mvlstm alone
if self.config["text1_attention"]:
q_w = Dense(1,
kernel_initializer=self.initializer_gate,
use_bias=False)(
q_embed) # use_bias=False to simple combination
show_layer_info('Dense', q_w)
q_w = Lambda(lambda x: softmax(x, axis=1),
output_shape=(self.config['text1_maxlen'], ),
name="q_w")(q_w)
show_layer_info('Lambda-softmax', q_w)
# ########## add attention weights for Q_words
q_w_layer = Lambda(lambda x: K.repeat_elements(
q_w, rep=self.config['embed_size'], axis=2))(q_w)
show_layer_info('repeat', q_w_layer)
q_embed = Multiply()([q_w_layer, q_embed])
show_layer_info('Dot-qw', q_embed)
# ####################### attention
d_embed = embedding(doc)
show_layer_info('Embedding_d', d_embed)
# ########## compute attention weights for the document words:
if self.config['text2_attention']:
d_w = Dense(1,
kernel_initializer=self.initializer_gate,
use_bias=False)(d_embed)
show_layer_info('Dense', d_w)
d_w = Lambda(lambda x: softmax(x, axis=1),
output_shape=(self.config['text2_maxlen'], ))(d_w)
show_layer_info('Lambda-softmax', d_w)
# ########## add attention weights for D_words
d_w_layer = Lambda(lambda x: K.repeat_elements(
d_w, rep=self.config['embed_size'], axis=2))(d_w)
d_embed = Multiply()([d_w_layer, d_embed])
show_layer_info('Dot-qw', d_embed)
# ####################### attention
q_rep = Bidirectional(
LSTM(self.config['hidden_size'],
return_sequences=True,
dropout=self.config['dropout_rate']))(q_embed)
show_layer_info('Bidirectional-LSTM_q', q_rep)
# ################# add attention for query positions:
if self.config["position_att_text1"]:
pos_w = Dense(1, activation='tanh')(
q_rep) # TimeDistributed(Dense(1, activation='tanh'))(q_rep)
pos_w = Flatten()(pos_w)
pos_w = Activation('softmax')(pos_w)
pos_w = RepeatVector(self.config['hidden_size'] * 2)(pos_w)
pos_w = Permute([2, 1])(pos_w)
q_rep = Multiply()([q_rep,
pos_w]) # merge([q_rep, pos_w], mode='mul')
d_rep = Bidirectional(
LSTM(self.config['hidden_size'],
return_sequences=True,
dropout=self.config['dropout_rate']))(d_embed)
show_layer_info('Bidirectional-LSTM_d', d_rep)
# ################# add attention for document positions:
if self.config["position_att_text2"]:
# https://machinelearningmastery.com/timedistributed-layer-for-long-short-term-memory-networks-in-python/
# timedistributed repeats the net between brackets for every input time step of the generated by the bi-LSTM
# so, same weights are applied for all the time steps. Without it, different weights are learned
pos_w = Dense(1, activation='tanh')(
d_rep) # TimeDistributed(Dense(1, activation='tanh'))(d_rep)
pos_w = Flatten()(pos_w)
pos_w = Activation('softmax')(pos_w)
pos_w = RepeatVector(self.config['hidden_size'] * 2)(pos_w)
pos_w = Permute([2, 1])(pos_w)
d_rep = Multiply()([d_rep,
pos_w]) # merge([d_rep, pos_w], mode='mul') #
cross = Match(match_type='dot')([q_rep, d_rep])
show_layer_info('Match-dot', cross)
cross_reshape = Reshape((-1, ))(cross)
show_layer_info('Reshape', cross_reshape)
mm_k = Lambda(lambda x: K.tf.nn.top_k(
x, k=self.config['topk'], sorted=True)[0])(cross_reshape)
show_layer_info('Lambda-topk', mm_k)
pool1_flat_drop = Dropout(rate=self.config['dropout_rate'])(mm_k)
show_layer_info('Dropout', pool1_flat_drop)
if self.config['target_mode'] == 'classification':
out_ = Dense(2, activation='softmax')(pool1_flat_drop)
elif self.config['target_mode'] in ['regression', 'ranking']:
out_ = Dense(1)(pool1_flat_drop)
show_layer_info('Dense', out_)
model = Model(inputs=[query, doc], outputs=out_)
plot_model(model,
to_file='../amvlstm.png',
show_shapes=True,
show_layer_names=True)
return model
def build(self):
query = Input(name='query', shape=(self.config['text1_maxlen'], ))
show_layer_info('Input', query)
doc = Input(name='doc', shape=(self.config['text2_maxlen'], ))
show_layer_info('Input', doc)
embedding = Embedding(self.config['vocab_size'],
self.config['embed_size'],
weights=[self.config['embed']],
trainable=self.embed_trainable)
q_embed = embedding(
Masking(mask_value=self.config['vocab_size'] - 1)(query))
show_layer_info('Embedding', q_embed)
d_embed = embedding(
Masking(mask_value=self.config['vocab_size'] - 1)(doc))
show_layer_info('Embedding', d_embed)
q_rep = Bidirectional(
LSTM(self.config['hidden_size'],
return_sequences=True,
dropout=self.config['dropout_rate']))(q_embed)
show_layer_info('Bidirectional-LSTM', q_rep)
d_rep = Bidirectional(
LSTM(self.config['hidden_size'],
return_sequences=True,
dropout=self.config['dropout_rate']))(d_embed)
show_layer_info('Bidirectional-LSTM', d_rep)
# Output size: sample,timestep,2*hidden_num
cross = Match(match_type=self.config['match_type'],
embedding_size=2 * self.config['hidden_size'],
r=5)([q_rep, d_rep])
# cross = Dot(axes=[2, 2])([q_embed, d_embed])
show_layer_info('Match', cross)
if self.config['match_type'] != 'tensor2':
cross_reshape = Reshape((-1, ))(cross)
show_layer_info('Reshape', cross_reshape)
mm_k = Lambda(lambda x: K.tf.nn.top_k(
x, k=self.config['topk'], sorted=True)[0])(cross_reshape)
show_layer_info('Lambda-topk', mm_k)
pool1_flat_drop = Dropout(rate=self.config['dropout_rate'])(mm_k)
show_layer_info('Dropout', pool1_flat_drop)
else:
act_cross = Activation('relu')(cross)
pool1_flat_drop = Lambda(lambda x: K.tf.reshape(
K.tf.nn.top_k(K.tf.transpose(
K.tf.reshape(x, (-1, x.shape[1] * x.shape[2], x.shape[3])),
[0, 2, 1]),
k=self.config['topk'],
sorted=True)[0],
(-1, K.tf.Dimension(self.config['topk'] * x.shape[3].value))))(
act_cross)
if self.config['target_mode'] == 'classification':
out_ = Dense(2, activation='softmax')(pool1_flat_drop)
elif self.config['target_mode'] in ['regression', 'ranking']:
out_ = Dense(1)(pool1_flat_drop)
show_layer_info('Dense', out_)
# model = Model(inputs=[query, doc, dpool_index], outputs=out_)
model = Model(inputs=[query, doc], outputs=out_)
return model