def build(self):
inputs = Input(shape=self.input_shape, name='encoder_input')
x = Dense(self.intermediate_dim,
activation=self.activation_fct)(inputs)
z_mean = Dense(self.latent_dim, name='z_mean')(x)
z_log_var = Dense(self.latent_dim, name='z_log_var')(x)
# use reparameterization trick to push the sampling out as input
# note that "output_shape" isn't necessary with the TensorFlow backend
z = Lambda(sampling, output_shape=(self.latent_dim, ),
name='z')([z_mean, z_log_var])
# instantiate encoder model
encoder = Model(inputs, [z_mean, z_log_var, z], name='encoder')
# build decoder model
latent_inputs = Input(shape=(self.latent_dim, ), name='z_sampling')
x = Dense(self.intermediate_dim,
activation=self.activation_fct)(latent_inputs)
outputs = Dense(self.original_dim, activation='sigmoid')(x)
# instantiate decoder model
decoder = Model(latent_inputs, outputs, name='decoder')
# instantiate VAE model
outputs = decoder(encoder(inputs)[2])
vae = Model(inputs, outputs, name='vae_mlp')
# VAE Loss = mse_loss or xent_loss + kl_loss
reconstruction_loss = mse(inputs, outputs)
reconstruction_loss *= self.original_dim
kl_loss = 1 + z_log_var - K.square(z_mean) - K.exp(z_log_var)
kl_loss = K.sum(kl_loss, axis=-1)
kl_loss *= -0.5
vae_loss = K.mean(reconstruction_loss + kl_loss)
vae.add_loss(vae_loss)
vae.compile(optimizer=self.optimizer,
loss=self.loss,
metrics=['accuracy'])
x_train_split, x_valid_split = train_test_split(
self.x_train,
test_size=self.train_test_split,
random_state=self.seed)
vae.fit(x_train_split,
x_train_split,
batch_size=self.batch_size,
epochs=self.epochs,
verbose=self.verbosity,
shuffle=True,
validation_data=(x_valid_split, x_valid_split))
x_train_pred = vae.predict(self.x_train)
train_mse = np.mean(np.power(self.x_train - x_train_pred, 2), axis=1)
self.threshold = np.quantile(train_mse, 0.9)
self.vae = vae
def build(self):
input_img = Input(shape=(28, 28, 1))
cnn = Conv2D(32, (3, 3), activation='relu', padding='same')(input_img)
cnn = MaxPooling2D((2, 2), padding='same')(cnn)
cnn = Conv2D(32, (3, 3), activation='relu', padding='same')(cnn)
cnn = MaxPooling2D((2, 2), padding='same')(cnn)
cnn = Conv2D(32, (3, 3), activation='relu', padding='same')(cnn)
encoded = MaxPooling2D((2, 2), padding='same')(cnn)
cnn = Conv2D(32, (3, 3), activation='relu', padding='same')(encoded)
cnn = UpSampling2D((2, 2))(cnn)
cnn = Conv2D(32, (3, 3), activation='relu', padding='same')(cnn)
cnn = UpSampling2D((2, 2))(cnn)
cnn = Conv2D(32, (3, 3), activation='relu')(cnn)
cnn = UpSampling2D((2, 2))(cnn)
decoded = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(cnn)
cnn_autoencoder = Model(input_img, decoded)
cnn_autoencoder.compile(optimizer='adam', loss='binary_crossentropy')
x_train = self.x_train.reshape(-1, 28, 28, 1)
x_train_split, x_valid_split = train_test_split(x_train, test_size=self.train_test_split,
random_state=self.seed)
cnn_autoencoder.fit(x_train_split, x_train_split,
epochs=self.epochs,
batch_size=self.batch_size,
validation_data=(x_valid_split, x_valid_split),
verbose=self.verbosity)
x_train_pred = cnn_autoencoder.predict(x_train)
mse = np.mean(np.power(x_train - x_train_pred, 2), axis=1)
# Semi-supervised due to given threshold
self.threshold = np.quantile(mse, 0.9)
self.cnn_autoencoder = cnn_autoencoder
class Autoencoder(object):
"""docstring for Autoencoder"""
# def __init__(self, sample_weights, sample_weight_mode):
def __init__(self, epochs, verbosity):
self.epochs = epochs
self.batch_size = 256
self.shuffle = True
self.validation_split = 0.05
self.optimizer = 'adadelta'
self.loss = 'mse'
self.verbosity = verbosity
self.code_layer_type = None
self.model = None
self.sample_weight_mode = None
self.sample_weights = None
self.y_true = None
self.y_pred = None
def model(self, code_layer_type, input_dim, code_dim):
self.code_layer_type = code_layer_type
assert len(code_dim) > 0
if self.code_layer_type == 'lstm':
assert len(input_dim) == 2
input_data = Input(shape=(input_dim[0], input_dim[1]))
if len(code_dim) == 1:
encoded = LSTM(code_dim[0])(input_data)
decoded = RepeatVector(input_dim[0])(encoded)
elif len(code_dim) > 1:
encoded = input_data
for i, units in enumerate(code_dim):
if i == len(code_dim) - 1:
encoded = LSTM(units)(encoded)
continue
encoded = LSTM(units, return_sequences=True)(encoded)
for i, units in enumerate(reversed(code_dim)):
if i == 1:
decoded = LSTM(units, return_sequences=True)(
RepeatVector(input_dim[0])(encoded))
elif i > 1:
decoded = LSTM(units, return_sequences=True)(decoded)
else:
raise ValueError("The codDim must be over 0.")
decoded = LSTM(input_dim[-1], return_sequences=True)(decoded)
self.model = Model(input_data, decoded)
elif self.code_layer_type == 'dense':
assert len(input_dim) == 1
input_data = Input(shape=(input_dim[0], ))
encoded = input_data
for i, units in enumerate(code_dim):
encoded = Dense(units, activation='relu')(encoded)
decoded = Dense(input_dim[-1], activation='sigmoid')(encoded)
self.model = Model(input_data, decoded)
elif self.code_layer_type == 'cov':
pass
def modelMasking(self, code_layer_type, input_dim, code_dim):
self.code_layer_type = code_layer_type
assert len(code_dim) > 0
if self.code_layer_type == 'lstm':
assert len(input_dim) == 2
input_data = Input(shape=(input_dim[0], input_dim[1]))
mask = Masking(mask_value=0.)(input_data)
if len(code_dim) == 1:
encoded = LSTM(code_dim[0])(mask)
decoded = RepeatVector(input_dim[0])(encoded)
elif len(code_dim) > 1:
encoded = mask
for i, units in enumerate(code_dim):
if i == len(code_dim) - 1:
encoded = LSTM(units)(encoded)
continue
encoded = LSTM(units, return_sequences=True)(encoded)
for i, units in enumerate(reversed(code_dim)):
if i == 1:
decoded = LSTM(units, return_sequences=True)(
RepeatVector(input_dim[0])(encoded))
elif i > 1:
decoded = LSTM(units, return_sequences=True)(decoded)
else:
raise ValueError("The codDim must be over 0.")
decoded = LSTM(input_dim[-1], return_sequences=True)(decoded)
self.model = Model(input_data, decoded)
elif self.code_layer_type == 'cov':
pass
elif self.code_layer_type == 'dense':
assert len(input_dim) == 1
input_data = Input(shape=(input_dim[0], ))
# encoded = input_data
# for i, units in enumerate(codeDim):
# encoded = Dense(units, activation='relu')(encoded)
# decoded = Dense(inputDim[-1], activation='sigmoid')(encoded)
# self.model = Model(input_data, decoded)
encoder = Dense(
code_dim[0],
activation="tanh",
activity_regularizer=regularizers.l1(10e-5))(input_data)
encoder = Dense(int(code_dim[0] / 2), activation="relu")(encoder)
decoder = Dense(int(code_dim[0] / 2), activation='tanh')(encoder)
decoder = Dense(input_dim[0], activation='relu')(decoder)
self.model = Model(input_data, decoder)
def compile(self, *args):
if len(args) == 0:
self.model.compile(optimizer=self.optimizer, loss=self.loss)
elif len(args) == 1:
if args[0] == 'temporal':
self.sample_weight_mode = args[0]
self.model.compile(optimizer=self.optimizer,
loss=self.loss,
sample_weight_mode=self.sample_weight_mode)
elif args[0] == 'customFunction':
self.model.compile(optimizer=self.optimizer,
loss=self.weighted_vector_mse)
else:
raise ValueError(
"Invalid maskType, please input 'sample_weights' or 'customFunction'"
)
else:
raise ValueError("argument # must be 0 or 1.")
def fit(self, *args):
# early_stopping = EarlyStopping(monitor='val_loss', min_delta=0.01, patience=3, verbose=1, mode='auto')
if len(args) == 2:
if args[1] == 'nor':
self.model.fit(args[0],
args[0],
epochs=self.epochs,
batch_size=self.batch_size,
shuffle=self.shuffle,
validation_split=self.validation_split,
verbose=self.verbosity)
# callbacks = [early_stopping])
elif args[1] == 'rev':
self.model.fit(args[0],
np.flip(args[0], 1),
epochs=self.epochs,
batch_size=self.batch_size,
shuffle=self.shuffle,
validation_split=self.validation_split,
verbose=self.verbosity)
# callbacks=[early_stopping])
else:
raise ValueError(
"decoding sequence type: 'normal' or 'reverse'.")
elif len(args) == 3:
self.sample_weights = args[2]
if args[1] == 'nor':
self.model.fit(args[0],
args[0],
epochs=self.epochs,
batch_size=self.batch_size,
shuffle=self.shuffle,
validation_split=self.validation_split,
sample_weight=self.sample_weights,
verbose=self.verbosity)
# callbacks=[early_stopping])
elif args[1] == 'rev':
self.model.fit(args[0],
np.flip(args[0], 1),
epochs=self.epochs,
batch_size=self.batch_size,
shuffle=self.shuffle,
validation_split=self.validation_split,
sample_weight=self.sample_weights,
verbose=self.verbosity)
# callbacks=[early_stopping])
else:
raise ValueError(
"Please input, 'data', 'nor' or 'rev', 'sample_weights'")
def predict(self, data):
return self.model.predict(data)
def weighted_vector_mse(self, y_true, y_pred):
self.y_true = y_true
self.y_pred = y_pred
weight = T.ceil(self.y_true)
loss = T.square(weight * (self.y_true - self.y_pred))
# use appropriate relations for other objectives. E.g, for binary_crossentropy:
# loss = weights * (y_true * T.log(y_pred) + (1.0 - y_true) * T.log(1.0 - y_pred))
return T.mean(T.sum(loss, axis=-1))
class QueryReformulation:
def __init__(self, model_path=None, output_path=''):
self.model = None
self.model_name = None
self.model_output = output_path + '/qr_{name}_model_[e{epoch}]_[p{precision}]_' \
+ str(datetime.now().date()) + '.h5'
if model_path:
self.model = load_model(model_path)
self.model.summary()
def build_model(self, model_name, query_dim, terms_dim, output_dim,
word_embedding):
self.model_name = model_name
query_input = Input(shape=(query_dim, ), name='query_input')
terms_input = Input(shape=(terms_dim, ), name='terms_input')
if model_name == 'lstm':
embedding_feature_block = Sequential(layers=[
Embedding(word_embedding.vocabulary_size,
word_embedding.dimensions,
weights=[word_embedding.embedding_matrix],
trainable=True,
mask_zero=False),
BatchNormalization(),
LSTM(64, return_sequences=True)
])
elif model_name == 'bilstm':
embedding_feature_block = Sequential(layers=[
Embedding(word_embedding.vocabulary_size,
word_embedding.dimensions,
weights=[word_embedding.embedding_matrix],
trainable=True,
mask_zero=False),
BatchNormalization(),
Bidirectional(LSTM(64, return_sequences=True))
])
else: # default cnn
embedding_feature_block = Sequential(layers=[
Embedding(word_embedding.vocabulary_size,
word_embedding.dimensions,
weights=[word_embedding.embedding_matrix],
trainable=True,
mask_zero=False),
BatchNormalization(),
Conv1D(filters=64, kernel_size=3, strides=1),
MaxPooling1D(pool_size=3)
])
# Features
query_feature = embedding_feature_block(query_input)
terms_feature = embedding_feature_block(terms_input)
# Query-Terms alignment
attention = Dot(axes=-1)([query_feature, terms_feature])
softmax_attention = Lambda(lambda x: softmax(x, axis=1),
output_shape=unchanged_shape)(attention)
terms_aligned = Dot(axes=1)([softmax_attention, terms_feature])
# Aligned features
if model_name == 'lstm':
flatten_layer = LSTM(128, return_sequences=False)(terms_aligned)
elif model_name == 'bilstm':
flatten_layer = Bidirectional(LSTM(
128, return_sequences=False))(terms_aligned)
else: # default cnn
merged_cnn = Conv1D(filters=128, kernel_size=3,
strides=1)(terms_aligned)
merged_cnn = MaxPooling1D(pool_size=3)(merged_cnn)
flatten_layer = Flatten()(merged_cnn)
# Output
dense = BatchNormalization()(flatten_layer)
dense = Dense(64, activation='sigmoid')(dense)
out = Dense(output_dim, activation='linear')(dense)
self.model = Model(inputs=[query_input, terms_input], outputs=out)
self.model.compile(optimizer='adam', loss=losses.mean_squared_error)
self.model.summary()
def train_model(self,
query_objs,
query_sequence,
terms_sequence,
candidate_terms,
epochs=20,
batch_size=4):
best_precision = 0
pool = Pool(batch_size)
for e in range(epochs):
print('Epochs: %3d/%d' % (e + 1, epochs))
reward = np.zeros(shape=(len(query_objs)))
precision = np.zeros(shape=(len(query_objs)))
for i, query, q_seq, t_seq, terms in get_batch_data(
query_objs, query_sequence, terms_sequence,
candidate_terms, batch_size):
print(' [%4d-%-4d/%d]' % (i, i + batch_size, len(query_objs)))
weights = self.model.predict(x=[q_seq, t_seq])
batch_reward_precision = pool.map(evaluate_reward_precision,
zip(weights, terms, query))
batch_reward_precision = np.array(batch_reward_precision)
batch_reward = 0.8 * np.asarray(
batch_reward_precision[:,
0]) + 0.2 * reward[i:i + batch_size]
self.model.train_on_batch(x=[q_seq, t_seq],
y=weights,
sample_weight=batch_reward)
reward[i:i + batch_size] = batch_reward_precision[:, 0]
precision[i:i + batch_size] = batch_reward_precision[:, 1]
# Save model
avg_precision = precision.mean()
print(' Average precision %.5f on epoch %d, best precision %.5f' %
(avg_precision, e + 1, best_precision))
if avg_precision > best_precision:
best_precision = avg_precision
self.model.save(filepath=self.model_output.format(
name=self.model_name,
epoch=e + 1,
precision=round(avg_precision, 4)))
pool.close()
pool.join()
def test_model(self,
query_objs,
query_sequence,
terms_sequence,
candidate_terms,
batch_size=4):
pool = Pool(batch_size)
precision_recall = np.zeros(shape=(len(query_objs), 2))
for i, query, q_seq, t_seq, terms in get_batch_data(
query_objs, query_sequence, terms_sequence, candidate_terms,
batch_size):
print('[%4d-%-4d/%d]' % (i, i + batch_size, len(query_objs)))
weights = self.model.predict(x=[q_seq, t_seq])
batch_precision_recall = pool.map(evaluate_precision_recall,
zip(weights, terms, query))
precision_recall[i:i +
batch_size] = np.array(batch_precision_recall)
pool.close()
pool.join()
return precision_recall.mean(axis=0)
def reformulate_query(self,
query_sequence,
terms_sequence,
candidate_terms,
threshold=0.5):
weights = self.model.predict(x=[[query_sequence], [terms_sequence]])
reformulated_query = recreate_query(terms=candidate_terms,
weights=weights[0],
threshold=threshold)
return reformulated_query
