def cnn_att_adv(is_training):
if is_training:
with tf.variable_scope('cnn_att_adv', reuse=False):
framework = Framework(is_training=True)
word_embedding = framework.embedding.word_embedding()
pos_embedding = framework.embedding.pos_embedding()
embedding = framework.embedding.concat_embedding(
word_embedding, pos_embedding)
x = framework.encoder.cnn(embedding,
FLAGS.hidden_size,
framework.mask,
activation=tf.nn.relu)
logit, repre = framework.selector.attention(
x, framework.scope, framework.label_for_select)
# Add perturbation
loss = framework.classifier.softmax_cross_entropy(logit)
new_word_embedding = framework.adversarial(loss, word_embedding)
new_embedding = framework.embedding.concat_embedding(
new_word_embedding, pos_embedding)
# Train
with tf.variable_scope('cnn_att_adv', reuse=True):
x = framework.encoder.cnn(new_embedding,
FLAGS.hidden_size,
framework.mask,
activation=tf.nn.relu)
logit, repre = framework.selector.attention(
x, framework.scope, framework.label_for_select)
loss = framework.classifier.softmax_cross_entropy(logit)
output = framework.classifier.output(logit)
framework.init_train_model(loss,
output,
optimizer=tf.train.GradientDescentOptimizer)
framework.load_train_data()
framework.train()
else:
with tf.variable_scope('cnn_att_adv', reuse=False):
framework = Framework(is_training=False)
word_embedding = framework.embedding.word_embedding()
pos_embedding = framework.embedding.pos_embedding()
embedding = framework.embedding.concat_embedding(
word_embedding, pos_embedding)
x = framework.encoder.cnn(embedding,
FLAGS.hidden_size,
framework.mask,
activation=tf.nn.relu)
logit, repre = framework.selector.attention(
x, framework.scope, framework.label_for_select)
framework.init_test_model(logit)
framework.load_test_data()
framework.test()
def birnn_ave_adv(is_training):
if is_training:
with tf.variable_scope('birnn_ave_adv', reuse=False):
framework = Framework(is_training=True)
word_embedding = framework.embedding.word_embedding()
pos_embedding = framework.embedding.pos_embedding()
embedding = framework.embedding.concat_embedding(
word_embedding, pos_embedding)
x = framework.encoder.birnn(embedding, FLAGS.hidden_size,
framework.length)
logit, repre = framework.selector.average(x, framework.scope)
# Add perturbation
loss = framework.classifier.softmax_cross_entropy(logit)
new_word_embedding = framework.adversarial(loss, word_embedding)
new_embedding = framework.embedding.concat_embedding(
new_word_embedding, pos_embedding)
# Train
with tf.variable_scope('birnn_ave_adv', reuse=True):
x = framework.encoder.birnn(new_embedding, FLAGS.hidden_size,
framework.length)
logit, repre = framework.selector.average(x, framework.scope)
loss = framework.classifier.softmax_cross_entropy(logit)
output = output(logit)
framework.init_train_model(loss,
output,
optimizer=tf.train.GradientDescentOptimizer)
framework.load_train_data()
framework.train()
else:
with tf.variable_scope('birnn_ave_adv', reuse=False):
framework = Framework(is_training=False)
word_embedding = framework.embedding.word_embedding()
pos_embedding = framework.embedding.pos_embedding()
embedding = framework.embedding.concat_embedding(
word_embedding, pos_embedding)
x = framework.encoder.birnn(embedding, FLAGS.hidden_size,
framework.length)
logit, repre = framework.selector.average(x, framework.scope)
framework.init_test_model(tf.nn.softmax(logit))
framework.load_test_data()
framework.test()
def rnn_att_adv(is_training):
if is_training:
with tf.variable_scope('rnn_att_adv', reuse=False):
framework = Framework(is_training=True)
word_embedding = framework.embedding.word_embedding()
pos_embedding = framework.embedding.pos_embedding()
embedding = framework.embedding.concat_embedding(
word_embedding, pos_embedding)
x = framework.encoder.rnn(embedding)
x = framework.selector.attention(x)
# Add perturbation
loss = framework.classifier.softmax_cross_entropy(x)
new_word_embedding = framework.adversarial(loss, word_embedding)
new_embedding = framework.embedding.concat_embedding(
new_word_embedding, pos_embedding)
# Train
with tf.variable_scope('rnn_att_adv', reuse=True):
x = framework.encoder.rnn(new_embedding)
x = framework.selector.attention(x)
loss = framework.classifier.softmax_cross_entropy(x)
output = framework.classifier.output(x)
framework.init_train_model(loss,
output,
optimizer=tf.train.GradientDescentOptimizer)
framework.load_train_data()
framework.train()
else:
with tf.variable_scope('rnn_att_adv', reuse=False):
framework = Framework(is_training=False)
word_embedding = framework.embedding.word_embedding()
pos_embedding = framework.embedding.pos_embedding()
embedding = framework.embedding.concat_embedding(
word_embedding, pos_embedding)
x = framework.encoder.rnn(embedding)
x = framework.selector.attention(x)
framework.init_test_model(tf.nn.softmax(x))
framework.load_test_data()
framework.test()
