def train_oracle(self):
self.init_oracle_trainng()
self.sess.run(tf.global_variables_initializer())
self.pre_epoch_num = 80
self.adversarial_epoch_num = 100
self.log = open('experiment-log-pgbleu3.csv', 'w')
generate_samples(self.sess, self.oracle, self.batch_size,
self.generate_num, self.oracle_file)
generate_samples(self.sess, self.generator, self.batch_size,
self.generate_num, self.generator_file)
self.gen_data_loader.create_batches(self.oracle_file)
self.oracle_data_loader.create_batches(self.generator_file)
self.init_metric()
# rollout = Reward(generator, update_rate)
print('start pre-train generator:')
for epoch in range(self.pre_epoch_num):
start = time()
loss = pre_train_epoch(self.sess, self.generator,
self.gen_data_loader)
end = time()
print('epoch:' + str(self.epoch) + '\t time:' + str(end - start))
self.add_epoch()
if epoch % 5 == 0:
self.evaluate()
self.reset_epoch()
print('start pg-bleu training:')
self.reward = Reward(self.oracle_file)
for epoch in range(self.adversarial_epoch_num):
start = time()
print('epoch:' + str(epoch))
for index in range(10):
samples = self.generator.generate(self.sess)
rewards = self.reward.get_reward(samples)
feed = {
self.generator.x: samples,
self.generator.rewards: rewards
}
_ = self.sess.run(self.generator.g_updates, feed_dict=feed)
end = time()
self.add_epoch()
print('epoch:' + str(self.epoch) + '\t time:' + str(end - start))
if epoch % 5 == 0 or epoch == self.adversarial_epoch_num - 1:
self.evaluate()
def train_oracle(self):
self.init_oracle_trainng()
self.sess.run(tf.compat.v1.global_variables_initializer())
self.log = open(self.log_file, 'w')
generate_samples(self.sess, self.oracle, self.batch_size,
self.generate_num, self.oracle_file)
generate_samples(self.sess, self.generator, self.batch_size,
self.generate_num, self.generator_file)
self.gen_data_loader.create_batches(self.oracle_file)
self.oracle_data_loader.create_batches(self.generator_file)
self.init_metric()
print('Pre-training Generator...')
for epoch in range(self.pre_epoch_num):
start = time()
loss = pre_train_epoch(self.sess, self.generator,
self.gen_data_loader)
end = time()
self.add_epoch()
if epoch % 5 == 0:
self.evaluate()
self.reset_epoch()
print('start pg-bleu training:')
self.reward = Reward(self.oracle_file)
for epoch in range(self.adversarial_epoch_num):
start = time()
for index in range(10):
samples = self.generator.generate(self.sess)
rewards = self.reward.get_reward(samples)
feed = {
self.generator.x: samples,
self.generator.rewards: rewards
}
_ = self.sess.run(self.generator.g_updates, feed_dict=feed)
end = time()
self.add_epoch()
if epoch % 5 == 0 or epoch == self.adversarial_epoch_num - 1:
self.evaluate()
class Pgbleu(Gan):
def __init__(self, oracle=None):
super().__init__()
# you can change parameters, generator here
self.vocab_size = 20
self.emb_dim = 32
self.hidden_dim = 32
self.sequence_length = 20
self.filter_size = [2, 3]
self.num_filters = [100, 200]
self.l2_reg_lambda = 0.2
self.dropout_keep_prob = 0.75
self.batch_size = 64
self.generate_num = 128
self.start_token = 0
self.oracle_file = 'save/oracle.txt'
self.generator_file = 'save/generator.txt'
def init_oracle_trainng(self, oracle=None):
if oracle is None:
oracle = OracleLstm(num_vocabulary=self.vocab_size,
batch_size=self.batch_size,
emb_dim=self.emb_dim,
hidden_dim=self.hidden_dim,
sequence_length=self.sequence_length,
start_token=self.start_token)
self.set_oracle(oracle)
generator = Generator(num_vocabulary=self.vocab_size,
batch_size=self.batch_size,
emb_dim=self.emb_dim,
hidden_dim=self.hidden_dim,
sequence_length=self.sequence_length,
start_token=self.start_token)
self.set_generator(generator)
gen_dataloader = DataLoader(batch_size=self.batch_size,
seq_length=self.sequence_length)
oracle_dataloader = DataLoader(batch_size=self.batch_size,
seq_length=self.sequence_length)
dis_dataloader = None
self.set_data_loader(gen_loader=gen_dataloader,
dis_loader=dis_dataloader,
oracle_loader=oracle_dataloader)
def init_metric(self):
nll = Nll(data_loader=self.oracle_data_loader,
rnn=self.oracle,
sess=self.sess)
self.add_metric(nll)
inll = Nll(data_loader=self.gen_data_loader,
rnn=self.generator,
sess=self.sess)
inll.set_name('nll-test')
self.add_metric(inll)
from utils.metrics.DocEmbSim import DocEmbSim
docsim = DocEmbSim(oracle_file=self.oracle_file,
generator_file=self.generator_file,
num_vocabulary=self.vocab_size)
self.add_metric(docsim)
def train_discriminator(self):
generate_samples(self.sess, self.generator, self.batch_size,
self.generate_num, self.generator_file)
self.dis_data_loader.load_train_data(self.oracle_file,
self.generator_file)
for _ in range(3):
self.dis_data_loader.next_batch()
x_batch, y_batch = self.dis_data_loader.next_batch()
feed = {
self.discriminator.input_x: x_batch,
self.discriminator.input_y: y_batch,
}
_ = self.sess.run(self.discriminator.train_op, feed)
def evaluate(self):
generate_samples(self.sess, self.generator, self.batch_size,
self.generate_num, self.generator_file)
self.oracle_data_loader.create_batches(self.generator_file)
if self.log is not None:
if self.epoch == 0 or self.epoch == 1:
for metric in self.metrics:
self.log.write(metric.get_name() + ',')
self.log.write('\n')
scores = super().evaluate()
for score in scores:
self.log.write(str(score) + ',')
self.log.write('\n')
return scores
return super().evaluate()
def train_oracle(self):
self.init_oracle_trainng()
self.sess.run(tf.global_variables_initializer())
self.pre_epoch_num = 80
self.adversarial_epoch_num = 100
self.log = open('experiment-log-pgbleu3.csv', 'w')
generate_samples(self.sess, self.oracle, self.batch_size,
self.generate_num, self.oracle_file)
generate_samples(self.sess, self.generator, self.batch_size,
self.generate_num, self.generator_file)
self.gen_data_loader.create_batches(self.oracle_file)
self.oracle_data_loader.create_batches(self.generator_file)
self.init_metric()
print('start pre-train generator:')
for epoch in range(self.pre_epoch_num):
start = time()
loss = pre_train_epoch(self.sess, self.generator,
self.gen_data_loader)
end = time()
print('epoch:' + str(self.epoch) + '\t time:' + str(end - start))
self.add_epoch()
if epoch % 5 == 0:
self.evaluate()
self.reset_epoch()
print('start pg-bleu training:')
self.reward = Reward(self.oracle_file)
for epoch in range(self.adversarial_epoch_num):
start = time()
print('epoch:' + str(epoch))
for index in range(10):
samples = self.generator.generate(self.sess)
rewards = self.reward.get_reward(samples)
feed = {
self.generator.x: samples,
self.generator.rewards: rewards
}
_ = self.sess.run(self.generator.g_updates, feed_dict=feed)
end = time()
self.add_epoch()
print('epoch:' + str(self.epoch) + '\t time:' + str(end - start))
if epoch % 5 == 0 or epoch == self.adversarial_epoch_num - 1:
self.evaluate()