def carnn(embedding_matrix,
config,
compare_out_size=CARNN_COMPARE_LAYER_OUTSIZE,
rnn_size=CARNN_RNN_SIZE,
rnn_dropout=CARNN_AGGREATION_DROPOUT):
q1 = Input(shape=(config['max_length'], ), dtype='int32', name='q1_input')
q2 = Input((config['max_length'], ), dtype='int32', name='q2_input')
activation = 'elu'
compare_dim = 500
compare_dropout = 0.2
embedding_layer = Embedding(embedding_matrix.shape[0],
embedding_matrix.shape[1],
trainable=config['embed_trainable'],
weights=[embedding_matrix]
# mask_zero=True
)
q1_embed = embedding_layer(q1)
q2_embed = embedding_layer(q2) # bsz, 1, emb_dims
q1_embed = BatchNormalization(axis=2)(q1_embed)
q2_embed = BatchNormalization(axis=2)(q2_embed)
q1_embed = SpatialDropout1D(config['spatial_dropout_rate'])(q1_embed)
q2_embed = SpatialDropout1D(config['spatial_dropout_rate'])(q2_embed)
highway_encoder = TimeDistributed(Highway(activation='relu'))
self_attention = SelfAttention(d_model=embedding_matrix.shape[1])
q1_encoded = highway_encoder(q1_embed, )
q2_encoded = highway_encoder(q2_embed, )
s1_encoded = self_attention(q1, q1_encoded)
s2_encoded = self_attention(q2, q2_encoded)
# Attention
q1_aligned, q2_aligned = soft_attention_alignment(q1_encoded, q2_encoded)
# Compare
q1_combined1 = Concatenate()([
q1_encoded,
q2_aligned,
interaction(q1_encoded, q2_aligned),
])
q1_combined2 = Concatenate()([
q2_aligned,
q1_encoded,
interaction(q1_encoded, q2_aligned),
])
q2_combined1 = Concatenate()([
q2_encoded,
q1_aligned,
interaction(q2_encoded, q1_aligned),
])
q2_combined2 = Concatenate()([
q1_aligned,
q2_encoded,
interaction(q2_encoded, q1_aligned),
])
s1_combined1 = Concatenate()([
q1_encoded,
s1_encoded,
interaction(q1_encoded, s1_encoded),
])
s1_combined2 = Concatenate()([
s1_encoded,
q1_encoded,
interaction(q1_encoded, s1_encoded),
])
s2_combined1 = Concatenate()([
q2_encoded,
s2_encoded,
interaction(q2_encoded, s2_encoded),
])
s2_combined2 = Concatenate()([
s2_encoded,
q2_encoded,
interaction(q2_encoded, s2_encoded),
])
compare_layers_d = [
Dense(compare_dim, activation=activation),
Dropout(compare_dropout),
Dense(compare_out_size, activation=activation),
Dropout(compare_dropout),
]
compare_layers_g = [
Dense(compare_dim, activation=activation),
Dropout(compare_dropout),
Dense(compare_out_size, activation=activation),
Dropout(compare_dropout),
]
# NOTE these can be optimized
q1_compare1 = time_distributed(q1_combined1, compare_layers_d)
q1_compare2 = time_distributed(q1_combined2, compare_layers_d)
q1_compare = Average()([q1_compare1, q1_compare2])
q2_compare1 = time_distributed(q2_combined1, compare_layers_d)
q2_compare2 = time_distributed(q2_combined2, compare_layers_d)
q2_compare = Average()([q2_compare1, q2_compare2])
s1_compare1 = time_distributed(s1_combined1, compare_layers_g)
s1_compare2 = time_distributed(s1_combined2, compare_layers_g)
s1_compare = Average()([s1_compare1, s1_compare2])
s2_compare1 = time_distributed(s2_combined1, compare_layers_g)
s2_compare2 = time_distributed(s2_combined2, compare_layers_g)
s2_compare = Average()([s2_compare1, s2_compare2])
# Aggregate
q1_encoded = Concatenate()([q1_encoded, q1_compare, s1_compare])
q2_encoded = Concatenate()([q2_encoded, q2_compare, s2_compare])
aggreate_rnn = CuDNNGRU(rnn_size, return_sequences=True)
q1_aggreated = aggreate_rnn(q1_encoded)
q1_aggreated = Dropout(rnn_dropout)(q1_aggreated)
q2_aggreated = aggreate_rnn(q2_encoded)
q2_aggreated = Dropout(rnn_dropout)(q2_aggreated)
# Pooling
q1_rep = apply_multiple(q1_aggreated, [
GlobalAvgPool1D(),
GlobalMaxPool1D(),
])
q2_rep = apply_multiple(q2_aggreated, [
GlobalAvgPool1D(),
GlobalMaxPool1D(),
])
q_diff = Lambda(lambda x: K.abs(x[0] - x[1]))([q1_rep, q2_rep])
q_multi = Lambda(lambda x: x[0] * x[1])([q1_rep, q2_rep])
feature_input = Input(shape=(config['feature_length'], ))
feature_dense = BatchNormalization()(feature_input)
feature_dense = Dense(config['dense_dim'],
activation='relu')(feature_dense)
h_all1 = Concatenate()([q1_rep, q2_rep, q_diff, q_multi, feature_dense])
h_all2 = Concatenate()([q2_rep, q1_rep, q_diff, q_multi, feature_dense])
h_all1 = Dropout(0.5)(h_all1)
h_all2 = Dropout(0.5)(h_all2)
dense = Dense(256, activation='relu')
h_all1 = dense(h_all1)
h_all2 = dense(h_all2)
h_all = Average()([h_all1, h_all2])
predictions = Dense(1, activation='sigmoid')(h_all)
model = Model(inputs=[q1, q2, feature_input], outputs=predictions)
opt = optimizers.get(config['optimizer'])
K.set_value(opt.lr, config['learning_rate'])
model.compile(optimizer=opt, loss='binary_crossentropy', metrics=[f1])
return model
def esim_word_char(embedding_matrix, char_embedding_matrix, config):
if config['rnn'] == 'gru' and config['gpu']:
word_encode = Bidirectional(
CuDNNGRU(config['rnn_output_size'], return_sequences=True))
word_compose = Bidirectional(
CuDNNGRU(config['rnn_output_size'], return_sequences=True))
char_encode = Bidirectional(
CuDNNGRU(config['rnn_output_size'], return_sequences=True))
char_compose = Bidirectional(
CuDNNGRU(config['rnn_output_size'], return_sequences=True))
else:
word_encode = Bidirectional(
CuDNNLSTM(config['rnn_output_size'], return_sequences=True))
word_compose = Bidirectional(
CuDNNLSTM(config['rnn_output_size'], return_sequences=True))
char_encode = Bidirectional(
CuDNNLSTM(config['rnn_output_size'], return_sequences=True))
char_compose = Bidirectional(
CuDNNLSTM(config['rnn_output_size'], return_sequences=True))
q1 = Input(shape=(config['max_length'], ), dtype='int32', name='q1_input')
q2 = Input((config['max_length'], ), dtype='int32', name='q2_input')
embedding_layer = Embedding(embedding_matrix.shape[0],
embedding_matrix.shape[1],
trainable=config['embed_trainable'],
weights=[embedding_matrix]
# mask_zero=True
)
q1_embed = embedding_layer(q1)
q2_embed = embedding_layer(q2) # bsz, 1, emb_dims
q1_embed = BatchNormalization(axis=2)(q1_embed)
q2_embed = BatchNormalization(axis=2)(q2_embed)
q1_embed = SpatialDropout1D(config['spatial_dropout_rate'])(q1_embed)
q2_embed = SpatialDropout1D(config['spatial_dropout_rate'])(q2_embed)
q1_encoded = word_encode(q1_embed)
q2_encoded = word_encode(q2_embed)
q1_aligned, q2_aligned = soft_attention_alignment(q1_encoded, q2_encoded)
q1_combined = Concatenate()(
[q1_encoded, q2_aligned,
submult(q1_encoded, q2_aligned)])
q2_combined = Concatenate()(
[q2_encoded, q1_aligned,
submult(q2_encoded, q1_aligned)])
# q1_combined = Dropout(self.config['dense_dropout'])(q1_combined)
# q2_combined = Dropout(self.config['dense_dropout'])(q2_combined)
q1_compare = word_compose(q1_combined)
q2_compare = word_compose(q2_combined)
# Aggregate
q1_rep = apply_multiple(q1_compare, [GlobalAvgPool1D(), GlobalMaxPool1D()])
q2_rep = apply_multiple(q2_compare, [GlobalAvgPool1D(), GlobalMaxPool1D()])
# Classifier
sub_rep = Lambda(lambda x: K.abs(x[0] - x[1]))([q1_rep, q2_rep])
mul_rep = Lambda(lambda x: x[0] * x[1])([q1_rep, q2_rep])
# Classifier
merged = Concatenate()([q1_rep, q2_rep, sub_rep, mul_rep])
q1_char = Input(shape=(config['char_max_length'], ),
dtype='int32',
name='q1_char_input')
q2_char = Input((config['char_max_length'], ),
dtype='int32',
name='q2_char_input')
char_embedding_layer = Embedding(char_embedding_matrix.shape[0],
char_embedding_matrix.shape[1],
trainable=config['embed_trainable'],
weights=[char_embedding_matrix]
# mask_zero=True
)
q1_embed_char = char_embedding_layer(q1_char)
q2_embed_char = char_embedding_layer(q2_char) # bsz, 1, emb_dims
q1_embed_char = BatchNormalization(axis=2)(q1_embed_char)
q2_embed_char = BatchNormalization(axis=2)(q2_embed_char)
q1_embed_char = SpatialDropout1D(
config['spatial_dropout_rate'])(q1_embed_char)
q2_embed_char = SpatialDropout1D(
config['spatial_dropout_rate'])(q2_embed_char)
q1_encoded_char = char_encode(q1_embed_char)
q2_encoded_char = char_encode(q2_embed_char)
q1_aligned_char, q2_aligned_char = soft_attention_alignment(
q1_encoded_char, q2_encoded_char)
q1_combined_char = Concatenate()([
q1_encoded_char, q2_aligned_char,
submult(q1_encoded_char, q2_aligned_char)
])
q2_combined_char = Concatenate()([
q2_encoded_char, q1_aligned_char,
submult(q2_encoded_char, q1_aligned_char)
])
# q1_combined = Dropout(self.config['dense_dropout'])(q1_combined)
# q2_combined = Dropout(self.config['dense_dropout'])(q2_combined)
q1_compare_char = char_compose(q1_combined_char)
q2_compare_char = char_compose(q2_combined_char)
# Aggregate
q1_rep_char = apply_multiple(
q1_compare_char,
[GlobalAvgPool1D(), GlobalMaxPool1D()])
q2_rep_char = apply_multiple(
q2_compare_char,
[GlobalAvgPool1D(), GlobalMaxPool1D()])
# Classifier
sub_rep_char = Lambda(lambda x: K.abs(x[0] - x[1]))(
[q1_rep_char, q2_rep_char])
mul_rep_char = Lambda(lambda x: x[0] * x[1])([q1_rep_char, q2_rep_char])
# Classifier
merged = Concatenate()([q1_rep, q2_rep, sub_rep, mul_rep])
merged_char = Concatenate()(
[q1_rep_char, q2_rep_char, sub_rep_char, mul_rep_char])
dense = BatchNormalization()(merged)
dense = Dense(config['dense_dim'], activation='elu')(dense)
dense_char = BatchNormalization()(merged_char)
dense_char = Dense(config['dense_dim'], activation='elu')(dense_char)
feature_input = Input(shape=(config['feature_length'], ))
feature_dense = BatchNormalization()(feature_input)
feature_dense = Dense(config['dense_dim'],
activation='relu')(feature_dense)
dense = Concatenate()([dense, dense_char, feature_dense])
dense = BatchNormalization()(dense)
dense = Dropout(config['dense_dropout'])(dense)
dense = Dense(config['dense_dim'], activation='elu')(dense)
dense = BatchNormalization()(dense)
dense = Dropout(config['dense_dropout'])(dense)
predictions = Dense(1, activation='sigmoid')(dense)
model = Model(inputs=[q1, q2, q1_char, q2_char, feature_input],
outputs=predictions)
opt = optimizers.get(config['optimizer'])
K.set_value(opt.lr, config['learning_rate'])
model.compile(optimizer=opt, loss='binary_crossentropy', metrics=[f1])
return model
def decom(embedding_matrix, config):
q1 = Input(shape=(config['max_length'], ), dtype='int32', name='q1_input')
q2 = Input((config['max_length'], ), dtype='int32', name='q2_input')
projection_hidden = 300
activation = 'elu'
projection_dropout = 0.2
projection_dim = 300
compare_dim = 500 # 300
compare_dropout = 0.2
embedding_layer = Embedding(embedding_matrix.shape[0],
embedding_matrix.shape[1],
trainable=config['embed_trainable'],
weights=[embedding_matrix]
# mask_zero=True
)
q1_embed = embedding_layer(q1)
q2_embed = embedding_layer(q2) # bsz, 1, emb_dims
q1_embed = BatchNormalization(axis=2)(q1_embed)
q2_embed = BatchNormalization(axis=2)(q2_embed)
q1_embed = SpatialDropout1D(config['spatial_dropout_rate'])(q1_embed)
q2_embed = SpatialDropout1D(config['spatial_dropout_rate'])(q2_embed)
highway_encoder = TimeDistributed(Highway(activation='relu'))
q1_encoded = highway_encoder(q1_embed, )
q2_encoded = highway_encoder(q2_embed, )
# Attention
q1_aligned, q2_aligned = soft_attention_alignment(q1_encoded, q2_encoded)
# Compare
q1_combined = Concatenate()(
[q1_encoded, q2_aligned,
interaction(q1_encoded, q2_aligned)])
q2_combined = Concatenate()(
[q2_encoded, q1_aligned,
interaction(q2_encoded, q1_aligned)])
compare_layers = [
Dense(compare_dim, activation=activation),
Dropout(compare_dropout),
Dense(compare_dim, activation=activation),
Dropout(compare_dropout),
]
q1_compare = time_distributed(q1_combined, compare_layers)
q2_compare = time_distributed(q2_combined, compare_layers)
# Aggregate
q1_rep = apply_multiple(q1_compare, [GlobalAvgPool1D(), GlobalMaxPool1D()])
q2_rep = apply_multiple(q2_compare, [GlobalAvgPool1D(), GlobalMaxPool1D()])
sub_rep = Lambda(lambda x: K.abs(x[0] - x[1]))([q1_rep, q2_rep])
mul_rep = Lambda(lambda x: x[0] * x[1])([q1_rep, q2_rep])
# Dense meta featues
# meta_densed = BatchNormalization()(meta_features)
# meta_densed = Highway(activation='relu')(meta_densed)
# meta_densed = Dropout(0.2)(meta_densed)
# Classifier
merged = Concatenate()([q1_rep, q2_rep, sub_rep, mul_rep])
dense = BatchNormalization()(merged)
dense = Dense(config['dense_dim'], activation='elu')(dense)
dense = BatchNormalization()(dense)
dense = Dropout(config['dense_dropout'])(dense)
dense = Dense(config['dense_dim'], activation='elu')(dense)
dense = BatchNormalization()(dense)
dense = Dropout(config['dense_dropout'])(dense)
predictions = Dense(1, activation='sigmoid')(dense)
model = Model(inputs=[q1, q2], outputs=predictions)
opt = optimizers.get(config['optimizer'])
K.set_value(opt.lr, config['learning_rate'])
model.compile(optimizer=opt, loss='binary_crossentropy', metrics=[f1])
return model