def main():
random.seed(SEED)
np.random.seed(SEED)
assert START_TOKEN == 0
gen_data_loader = Gen_Data_loader(BATCH_SIZE)
likelihood_data_loader = Likelihood_data_loader(BATCH_SIZE)
vocab_size = 5000
generator = get_trainable_model(vocab_size)
target_params = cPickle.load(open('save/target_params.pkl'))
target_lstm = TARGET_LSTM(vocab_size, 64, 32, 32, 20, 0, target_params)
config = tf.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 0.5
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer)
generate_samples(sess, target_lstm, 64, 10000, positive_file)
gen_data_loader.create_batches(positive_file)
log = open('log/experiment-log.txt', 'w')
# pre-train generator
print 'Start pre-training...'
log.write('pre-training...\n')
for epoch in xrange(PRE_EPOCH_NUM):
print 'pre-train epoch:', epoch
loss = pre_train_epoch(sess, generator, gen_data_loader)
if epoch % 5 == 0:
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
print 'pre-train epoch ', epoch, 'test_loss ', test_loss
buffer = str(epoch) + ' ' + str(test_loss) + '\n'
log.write(buffer)
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = 'After supervised-training:' + ' ' + str(test_loss) + '\n'
log.write(buffer)
log.close()
def main():
random.seed(SEED)
np.random.seed(SEED)
# assert START_TOKEN == 0
gen_data_loader = Gen_Data_loader(BATCH_SIZE)
likelihood_data_loader = Likelihood_data_loader(BATCH_SIZE)
vocab_size = 5000
best_score = 9.5
generator = get_trainable_model(vocab_size)
# target_lstm = TARGET_LSTM(vocab_size, 64, 32, 32, 20, 0)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
# generate_samples(sess, target_lstm, 64, 10000, positive_file)
positive_data = np.load(positive_file).tolist()
gen_data_loader.create_batches(positive_data)
log = open('log/ss_experiment-log.txt', 'w')
# pre-train generator
print 'Start scheduled sampling training...'
log.write('scheduled sampling training...\n')
curriculum_rate = .9
for epoch in xrange(EPOCH_NUM):
curriculum_rate = max(0.0, curriculum_rate - 0.002)
loss = pre_train_epoch(sess, generator, gen_data_loader, curriculum_rate)
# generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
# likelihood_data_loader.create_batches(eval_file)
# test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
# print 'pre-train epoch ', epoch, 'curriculum rate:', curriculum_rate, 'test_loss ', test_loss
# buffer = str(epoch) + ' ' + str(curriculum_rate) + ' ' + str(test_loss) + '\n'
buffer = str(epoch) + ' ' + str(curriculum_rate) + ' ' + str(loss) + '\n'
print(buffer)
log.write(buffer)
# if test_loss < best_score:
# best_score = test_loss
# print 'best score: ', test_loss
# generate_samples(sess, generator, BATCH_SIZE, 100000, eval_file)
# likelihood_data_loader.create_batches(eval_file)
# significance_test(sess, target_lstm, likelihood_data_loader, 'significance/schedule_sampling.txt')
log.close()
generate_samples(sess, generator, BATCH_SIZE, 100, final_trans_file)
def main():
random.seed(SEED)
np.random.seed(SEED)
stringGenerator = TextGenerator('../corpus/index2word.pickle',
'../corpus/word2index.pickle',
'../corpus/all.code')
assert START_TOKEN == 0
gen_data_loader = Gen_Data_loader(BATCH_SIZE)
likelihood_data_loader = Likelihood_data_loader(BATCH_SIZE)
#vocab_size = 5000
vocab_size = len(stringGenerator.index2Word)
generator = get_trainable_model(vocab_size)
target_params = cPickle.load(open('save/target_params.pkl'))
target_params[00] = np.random.rand(vocab_size, 32).astype(np.float32)
target_params[-2] = np.random.rand(32, vocab_size).astype(np.float32)
target_params[-1] = np.random.rand(vocab_size).astype(np.float32)
target_lstm = TARGET_LSTM(vocab_size, 64, 32, 32, 20, 0, target_params)
config = tf.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 0.5
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.initialize_all_variables())
#generate_samples(sess, target_lstm, 64, 10000, positive_file)
stringGenerator.saveSamplesToFile(20, 10000, positive_file)
gen_data_loader.create_batches(positive_file)
log = open('log/experiment-log.txt', 'w')
# pre-train generator
print 'Start pre-training...'
log.write('pre-training...\n')
for epoch in xrange(PRE_EPOCH_NUM):
print 'pre-train epoch:', epoch
loss = pre_train_epoch(sess, generator, gen_data_loader)
if epoch % 5 == 0:
#generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
stringGenerator.saveSamplesToFile(20, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
print 'pre-train epoch ', epoch, 'test_loss ', test_loss
buffer = str(epoch) + ' ' + str(test_loss) + '\n'
log.write(buffer)
#generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
stringGenerator.saveSamplesToFile(20, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = 'After supervised-training:' + ' ' + str(test_loss) + '\n'
log.write(buffer)
log.close()
def main():
random.seed(SEED)
np.random.seed(SEED)
assert START_TOKEN == 0
# load data (likelihood?)
gen_data_loader = Gen_Data_loader(BATCH_SIZE)
likelihood_data_loader = Likelihood_data_loader(BATCH_SIZE)
vocab_size = 68
# load generator with parameters
generator = get_trainable_model(vocab_size)
# target_params = cPickle.load(open('save/target_params.pkl'))
target_params = initialize_parameters(68)
target_lstm = TARGET_LSTM(vocab_size, BATCH_SIZE, EMB_DIM, HIDDEN_DIM, SEQ_LENGTH, START_TOKEN, target_params)
config = tf.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 0.5
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
# generating synthetic data which constitute to original data
# generate_samples(sess, target_lstm, 64, 100, positive_file)
gen_data_loader.create_batches(positive_file)
log = open('log/experiment-log.txt', 'w')
# pre-train generator
print 'Start pre-training...'
log.write('pre-training...\n')
for epoch in xrange(PRE_EPOCH_NUM):
print 'pre-train epoch:', epoch
loss = pre_train_epoch(sess, generator, gen_data_loader)
if epoch % 5 == 0:
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
print 'pre-train epoch ', epoch, 'test_loss ', test_loss
buffer = str(epoch) + ' ' + str(test_loss) + '\n'
log.write(buffer)
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
print 'final pre-train epoch ', 'test_loss ', test_loss
buffer = 'After supervised-training:' + ' ' + str(test_loss) + '\n'
log.write(buffer)
log.close()
def main():
random.seed(SEED)
np.random.seed(SEED)
assert START_TOKEN == 0
gen_data_loader = Gen_Data_loader(BATCH_SIZE)
likelihood_data_loader = Likelihood_data_loader(BATCH_SIZE)
vocab_size = 5000
best_score = 9.1
generator = get_trainable_model(vocab_size)
target_lstm = TARGET_LSTM(vocab_size, BATCH_SIZE, EMB_DIM, HIDDEN_DIM, SEQ_LENGTH, START_TOKEN)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer)
generate_samples(sess, target_lstm, 64, 10000, positive_file)
################################################################
gen_data_loader.create_batches(positive_file)
references = load_references(positive_file)
log = open('log/experiment-log.txt', 'w')
# pre-train generator
print 'Start pre-training...'
log.write('pre-training...\n')
for epoch in xrange(PRE_EPOCH_NUM):
print 'pre-train epoch:', epoch
loss = pre_train_epoch(sess, generator, gen_data_loader)
if epoch % 5 == 0:
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
print 'pre-train epoch ', epoch, 'test_loss ', test_loss
buffer = str(epoch) + ' ' + str(test_loss) + '\n'
log.write(buffer)
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = 'After pre-training:' + ' ' + str(test_loss) + '\n'
log.write(buffer)
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
significance_test(sess, target_lstm, likelihood_data_loader, 'significance/supervise.txt')
rollout = ROLLOUT(generator, references)
print '#########################################################################'
print 'Start Reinforcement Training Generator...'
log.write('Reinforcement Training...\n')
for total_batch in range(TOTAL_BATCH):
for it in range(TRAIN_ITER):
samples = generator.generate(sess)
print 'start calculating BLEU...'
rewards = rollout.get_reward(sess, samples, 1, (1.0 / 3, 1.0 / 3, 1.0 / 3))
feed = {generator.x: samples, generator.rewards: rewards}
_, g_loss = sess.run([generator.g_updates, generator.g_loss], feed_dict=feed)
if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = str(total_batch) + ' ' + str(test_loss) + '\n'
print 'total_batch: ', total_batch, 'test_loss: ', test_loss
log.write(buffer)
if test_loss < best_score:
best_score = test_loss
print 'best score: ', test_loss
significance_test(sess, target_lstm, likelihood_data_loader, 'significance/pg_bleu.txt')
rollout.update_params()
log.close()
def main():
random.seed(SEED)
np.random.seed(SEED)
assert START_TOKEN == 0
gen_data_loader = Gen_Data_loader(BATCH_SIZE)
likelihood_data_loader = Likelihood_data_loader(BATCH_SIZE)
vocab_size = 68
dis_data_loader = Dis_dataloader()
best_score = 1000
# load generator with parameters
generator = get_trainable_model(vocab_size)
target_params = initialize_parameters(vocab_size)
target_lstm = TARGET_LSTM(vocab_size, BATCH_SIZE, EMB_DIM, HIDDEN_DIM,
SEQ_LENGTH, START_TOKEN, target_params)
# CNNs
with tf.variable_scope('discriminator'):
cnn = TextCNN(sequence_length=SEQ_LENGTH,
num_classes=2,
vocab_size=vocab_size,
embedding_size=dis_embedding_dim,
filter_sizes=dis_filter_sizes,
num_filters=dis_num_filters,
l2_reg_lambda=dis_l2_reg_lambda)
cnn_params = [
param for param in tf.trainable_variables()
if 'discriminator' in param.name
]
# Define Discriminator Training procedure
dis_global_step = tf.Variable(0, name="global_step", trainable=False)
dis_optimizer = tf.train.AdamOptimizer(1e-4)
dis_grads_and_vars = dis_optimizer.compute_gradients(cnn.loss,
cnn_params,
aggregation_method=2)
dis_train_op = dis_optimizer.apply_gradients(dis_grads_and_vars,
global_step=dis_global_step)
config = tf.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 0.5
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
# generate_samples(sess, target_lstm, 64, 10000, positive_file)
gen_data_loader.create_batches(positive_file)
log = open(logpath, 'w')
# pre-train generator
print 'Start pre-training...'
log.write('pre-training...\n')
for epoch in xrange(PRE_EPOCH_NUM):
print 'pre-train epoch:', epoch
loss = pre_train_epoch(sess, generator, gen_data_loader)
if epoch % 5 == 0:
file_name = 'target_generate/pretrain_epoch' + str(epoch) + '.pkl'
generate_samples(sess, generator, BATCH_SIZE, generated_num,
file_name)
likelihood_data_loader.create_batches(file_name)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
print 'pre-train epoch ', epoch, 'test_loss ', test_loss
buffer = str(epoch) + ' ' + str(test_loss) + '\n'
log.write(buffer)
if epoch % 100 != 0:
os.remove(file_name)
file_name = 'target_generate/pretrain_finished.pkl'
generate_samples(sess, generator, BATCH_SIZE, generated_num, file_name)
likelihood_data_loader.create_batches(file_name)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = 'After pre-training:' + ' ' + str(test_loss) + '\n'
log.write(buffer)
file_name = 'target_generate/supervise.pkl'
generate_samples(sess, generator, BATCH_SIZE, generated_num, file_name)
likelihood_data_loader.create_batches(file_name)
significance_test(sess, target_lstm, likelihood_data_loader,
'significance/supervise.txt')
os.remove(file_name)
print 'Start training discriminator...'
for i in range(dis_alter_epoch):
print 'dis_alter_epoch : ' + str(i)
generate_samples(sess, generator, BATCH_SIZE, generated_num,
negative_file)
# train discriminator
dis_x_train, dis_y_train = dis_data_loader.load_train_data(
positive_file, negative_file)
dis_batches = dis_data_loader.batch_iter(zip(dis_x_train, dis_y_train),
dis_batch_size,
dis_num_epochs)
for batch in dis_batches:
try:
x_batch, y_batch = zip(*batch)
feed = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: dis_dropout_keep_prob
}
_, step = sess.run([dis_train_op, dis_global_step], feed)
except ValueError:
pass
rollout = ROLLOUT(generator, 0.8)
print '#########################################################################'
print 'Start Reinforcement Training Generator...'
log.write('Reinforcement Training...\n')
for total_batch in range(TOTAL_BATCH):
for it in range(TRAIN_ITER):
samples = generator.generate(sess)
rewards = rollout.get_reward(sess, samples, 16, cnn)
feed = {generator.x: samples, generator.rewards: rewards}
_, g_loss = sess.run([generator.g_updates, generator.g_loss],
feed_dict=feed)
if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
file_name = 'target_generate/reinforce_batch' + str(
total_batch) + '.pkl'
generate_samples(sess, generator, BATCH_SIZE, generated_num,
file_name)
likelihood_data_loader.create_batches(file_name)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = str(total_batch) + ' ' + str(test_loss) + '\n'
print 'total_batch: ', total_batch, 'test_loss: ', test_loss
log.write(buffer)
if total_batch % 50 != 0:
os.remove(file_name)
if test_loss < best_score:
best_score = test_loss
print 'best score: ', test_loss
significance_test(sess, target_lstm, likelihood_data_loader,
'significance/seqgan.txt')
rollout.update_params()
# generate for discriminator
print 'Start training discriminator'
for _ in range(5):
# for _ in range(2):
generate_samples(sess, generator, BATCH_SIZE, generated_num,
negative_file)
dis_x_train, dis_y_train = dis_data_loader.load_train_data(
positive_file, negative_file)
dis_batches = dis_data_loader.batch_iter(
zip(dis_x_train, dis_y_train), dis_batch_size, 3)
for batch in dis_batches:
try:
x_batch, y_batch = zip(*batch)
feed = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: dis_dropout_keep_prob
}
_, step = sess.run([dis_train_op, dis_global_step], feed)
except ValueError:
pass
log.close()
def main():
random.seed(SEED)
np.random.seed(SEED)
gen_data_loader = Gen_Data_loader(BATCH_SIZE)
likelihood_data_loader = Likelihood_data_loader(BATCH_SIZE)
vocab_size = 5000
dis_data_loader = Dis_dataloader()
best_score = 1000
generator = get_trainable_model(vocab_size)
# oracle model : target lstm
# target_params = cPickle.load(open('save/target_params.pkl'))
# target_lstm = TARGET_LSTM(vocab_size, 64, 32, 32, SEQ_LENGTH, 0, target_params)
with tf.variable_scope('discriminator'):
cnn = TextCNN(
sequence_length=SEQ_LENGTH,
num_classes=2,
vocab_size=vocab_size,
embedding_size=dis_embedding_dim,
filter_sizes=dis_filter_sizes,
num_filters=dis_num_filters,
l2_reg_lambda=dis_l2_reg_lambda)
cnn_params = [param for param in tf.trainable_variables() if 'discriminator' in param.name]
# Define Discriminator Training procedure
dis_global_step = tf.Variable(0, name="global_step", trainable=False)
dis_optimizer = tf.train.AdamOptimizer(1e-4)
dis_grads_and_vars = dis_optimizer.compute_gradients(cnn.loss, cnn_params, aggregation_method=2)
dis_train_op = dis_optimizer.apply_gradients(dis_grads_and_vars, global_step=dis_global_step)
config = tf.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 0.5
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
# czq
# generate real data
# generate_samples(sess, target_lstm, 64, 10000, positive_file)
# store real data for next step
positive_data = np.load(positive_file).tolist()
gen_data_loader.create_batches(positive_data)
log = open('log/seq_mle_experiment-log.txt', 'w')
# pre-train generator
print '#########################################################################'
print 'Start pre-training generator...'
log.write('pre-training...\n')
for epoch in xrange(PRE_EPOCH_NUM):
# print 'pre-train epoch:', epoch
loss = pre_train_epoch(sess, generator, gen_data_loader)
if epoch % 5 == 0:
# generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
# likelihood_data_loader.create_batches(eval_file)
# test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
# print 'pre-train epoch ', epoch, 'test_loss ', test_loss
# buffer = str(epoch) + ' ' + str(test_loss) + '\n'
buffer = 'pre-trained generator:' + str(epoch) + ' ' + str(loss)
print(buffer)
log.write(buffer + '\n')
# generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
# likelihood_data_loader.create_batches(eval_file)
# test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
# buffer = 'After pre-training:' + ' ' + str(test_loss) + '\n'
buffer = 'After pre-training:' + ' ' + str(loss)
print(buffer)
log.write(buffer + '\n')
# generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
# likelihood_data_loader.create_batches(eval_file)
# significance_test(sess, target_lstm, likelihood_data_loader, 'significance/supervise.txt')
# test purpose only
generate_samples(sess, generator, BATCH_SIZE, 100, final_trans_file_mle)
# exit(0)
print 'Start pre-training discriminator...'
for _ in range(dis_alter_epoch):
generate_samples(sess, generator, BATCH_SIZE, generated_num, negative_file)
# train discriminator
dis_x_train, dis_y_train = dis_data_loader.load_train_data(positive_file, negative_file)
dis_batches = dis_data_loader.batch_iter(
zip(dis_x_train, dis_y_train), dis_batch_size, dis_num_epochs
)
for batch in dis_batches:
try:
x_batch, y_batch = zip(*batch)
feed = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: dis_dropout_keep_prob
}
_, step = sess.run([dis_train_op, dis_global_step], feed)
except Exception as e:
# print str(e)
raise
loss = sess.run(cnn.loss, feed)
buffer = 'pre-train discriminator' + ' ' + str(loss)
print buffer
log.write(buffer + '\n')
rollout = ROLLOUT(generator, 0.8)
print('Before GAN')
print '#########################################################################'
print 'Start Reinforcement Training Generator...'
log.write('Reinforcement Training...\n')
# for tensorboard
# writer = tf.summary.FileWriter('./tb_logs', graph=tf.get_default_graph())
for total_batch in range(TOTAL_BATCH):
print 'progress', total_batch, '/', TOTAL_BATCH
for it in range(TRAIN_ITER):
samples = generator.generate(sess)
rewards = rollout.get_reward(sess, samples, 16, cnn)
feed = {generator.x: samples, generator.rewards: rewards}
_, g_loss, pre_loss = sess.run([generator.g_updates, generator.g_loss, generator.pretrain_loss],
feed_dict=feed)
buffer = 'G-step:' + str(TRAIN_ITER) + ':' + str(g_loss) + '|' + str(pre_loss)
log.write(buffer + '\n')
print(buffer)
# if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
# generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
# likelihood_data_loader.create_batches(eval_file)
# # test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
# # buffer = str(total_batch) + ' ' + str(test_loss) + '\n'
# # print 'total_batch: ', total_batch, 'test_loss: ', test_loss
# log.write(buffer)
# if test_loss < best_score:
# best_score = test_loss
# print 'best score: ', test_loss
# significance_test(sess, target_lstm, likelihood_data_loader, 'significance/seqgan.txt')
rollout.update_params()
# generate for discriminator
print('Start training discriminator')
log.write('training discriminator...\n')
for _ in range(5):
generate_samples(sess, generator, BATCH_SIZE, generated_num, negative_file)
dis_x_train, dis_y_train = dis_data_loader.load_train_data(positive_file, negative_file)
dis_batches = dis_data_loader.batch_iter(zip(dis_x_train, dis_y_train), dis_batch_size, 3)
for batch in dis_batches:
try:
x_batch, y_batch = zip(*batch)
feed = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: dis_dropout_keep_prob
}
_, step = sess.run([dis_train_op, dis_global_step], feed)
except ValueError:
pass
loss = sess.run(cnn.loss, feed)
buffer = 'discriminator' + ' ' + str(loss)
print buffer
log.write(buffer + '\n')
log.close()
# save the model
# saver = tf.train.Saver({"gen": generator})
# saver.save(sess, 'my-model')
# generate samples
generate_samples(sess, generator, BATCH_SIZE, 100, final_trans_file_seqgan)
def main():
random.seed(SEED)
np.random.seed(SEED)
assert START_TOKEN == 0
gen_data_loader = Gen_Data_loader(BATCH_SIZE)
likelihood_data_loader = Likelihood_data_loader(BATCH_SIZE)
vocab_size = 5000
best_score = 9.1
generator = get_trainable_model(vocab_size)
# target_lstm = TARGET_LSTM(vocab_size, BATCH_SIZE, EMB_DIM, HIDDEN_DIM, SEQ_LENGTH, START_TOKEN)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
# generate_samples(sess, target_lstm, 64, 10000, positive_file)
################################################################
positive_data = np.load(positive_file).tolist()
gen_data_loader.create_batches(positive_data)
references = load_references(positive_data)
log = open('log/pg_experiment-log.txt', 'w')
# pre-train generator
print 'Start pre-training...'
log.write('pre-training...\n')
for epoch in xrange(PRE_EPOCH_NUM):
print 'pre-train epoch:', epoch
loss = pre_train_epoch(sess, generator, gen_data_loader)
if epoch % 5 == 0:
# generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
# likelihood_data_loader.create_batches(eval_file)
# test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
# print 'pre-train epoch ', epoch, 'test_loss ', test_loss
# buffer = str(epoch) + ' ' + str(test_loss) + '\n'
# log.write(buffer)
print 'pre-train epoch ', epoch, 'loss ', loss
buffer = str(epoch) + ' ' + str(loss) + '\n'
log.write(buffer)
# generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
# likelihood_data_loader.create_batches(eval_file)
# test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
# buffer = 'After pre-training:' + ' ' + str(test_loss) + '\n'
print 'After pre-train epoch ', loss
buffer = str(loss) + '\n'
log.write(buffer)
# generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
# likelihood_data_loader.create_batches(eval_file)
# significance_test(sess, target_lstm, likelihood_data_loader, 'significance/supervise.txt')
rollout = ROLLOUT(generator, references)
print '#########################################################################'
print 'Start Reinforcement Training Generator...'
log.write('Reinforcement Training...\n')
for total_batch in range(TOTAL_BATCH):
for it in range(TRAIN_ITER):
samples = generator.generate(sess)
print 'start calculating BLEU...'
rewards = rollout.get_reward(sess, samples, 1, (1.0 / 3, 1.0 / 3, 1.0 / 3))
feed = {generator.x: samples, generator.rewards: rewards}
_, g_loss = sess.run([generator.g_updates, generator.g_loss], feed_dict=feed)
# if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
# generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
# likelihood_data_loader.create_batches(eval_file)
# test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
# buffer = str(total_batch) + ' ' + str(test_loss) + '\n'
# print 'total_batch: ', total_batch, 'test_loss: ', test_loss
# log.write(buffer)
# if test_loss < best_score:
# best_score = test_loss
# print 'best score: ', test_loss
# significance_test(sess, target_lstm, likelihood_data_loader, 'significance/pg_bleu.txt')
print('Current loss:' + str(total_batch) + ':' + str(g_loss))
rollout.update_params()
log.close()
generate_samples(sess, generator, BATCH_SIZE, 100, final_trans_file)
def main():
random.seed(SEED)
np.random.seed(SEED)
stringGenerator = TextGenerator('../corpus/index2word.pickle',
'../corpus/word2index.pickle',
'../corpus/all.code')
assert START_TOKEN == 0
gen_data_loader = Gen_Data_loader(BATCH_SIZE)
likelihood_data_loader = Likelihood_data_loader(BATCH_SIZE)
vocab_size = len(stringGenerator.index2Word)
dis_data_loader = Dis_dataloader()
best_score = 1000
generator = get_trainable_model(vocab_size)
target_params = cPickle.load(open('save/target_params.pkl'))
target_params[00] = np.random.rand(vocab_size, 32).astype(np.float32)
target_params[-2] = np.random.rand(32, vocab_size).astype(np.float32)
target_params[-1] = np.random.rand(vocab_size).astype(np.float32)
target_lstm = TARGET_LSTM(vocab_size, 64, 32, 32, 20, 0, target_params)
with tf.variable_scope('discriminator'):
cnn = TextCNN(sequence_length=20,
num_classes=2,
vocab_size=vocab_size,
embedding_size=dis_embedding_dim,
filter_sizes=dis_filter_sizes,
num_filters=dis_num_filters,
l2_reg_lambda=dis_l2_reg_lambda)
cnn_params = [
param for param in tf.trainable_variables()
if 'discriminator' in param.name
]
# Define Discriminator Training procedure
dis_global_step = tf.Variable(0, name="global_step", trainable=False)
dis_optimizer = tf.train.AdamOptimizer(1e-4)
dis_grads_and_vars = dis_optimizer.compute_gradients(cnn.loss,
cnn_params,
aggregation_method=2)
dis_train_op = dis_optimizer.apply_gradients(dis_grads_and_vars,
global_step=dis_global_step)
config = tf.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 0.5
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.initialize_all_variables())
#generate_samples(sess, target_lstm, 64, 10000, positive_file)
stringGenerator.saveSamplesToFile(20, 10000, positive_file)
gen_data_loader.create_batches(positive_file)
log = open('log/experiment-log.txt', 'w')
# pre-train generator
print 'Start pre-training...'
log.write('pre-training...\n')
for epoch in xrange(PRE_EPOCH_NUM):
print 'pre-train epoch:', epoch
loss = pre_train_epoch(sess, generator, gen_data_loader)
if epoch % 5 == 0:
generate_samples(sess, generator, BATCH_SIZE, generated_num,
eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
print 'pre-train epoch ', epoch, 'test_loss ', test_loss
buffer = str(epoch) + ' ' + str(test_loss) + '\n'
log.write(buffer)
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = 'After pre-training:' + ' ' + str(test_loss) + '\n'
log.write(buffer)
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
significance_test(sess, target_lstm, likelihood_data_loader,
'significance/supervise.txt')
print 'Start training discriminator...'
for _ in range(dis_alter_epoch):
generate_samples(sess, generator, BATCH_SIZE, generated_num,
negative_file)
# train discriminator
dis_x_train, dis_y_train = dis_data_loader.load_train_data(
positive_file, negative_file)
dis_batches = dis_data_loader.batch_iter(zip(dis_x_train, dis_y_train),
dis_batch_size,
dis_num_epochs)
for batch in dis_batches:
try:
x_batch, y_batch = zip(*batch)
feed = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: dis_dropout_keep_prob
}
_, step = sess.run([dis_train_op, dis_global_step], feed)
except ValueError:
pass
rollout = ROLLOUT(generator, 0.8)
print '#########################################################################'
print 'Start Reinforcement Training Generator...'
log.write('Reinforcement Training...\n')
for total_batch in range(TOTAL_BATCH):
for it in range(TRAIN_ITER):
samples = generator.generate(sess)
rewards = rollout.get_reward(sess, samples, 16, cnn)
feed = {generator.x: samples, generator.rewards: rewards}
_, g_loss = sess.run([generator.g_updates, generator.g_loss],
feed_dict=feed)
if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
generate_samples(sess, generator, BATCH_SIZE, generated_num,
eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = str(total_batch) + ' ' + str(test_loss) + '\n'
print 'total_batch: ', total_batch, 'test_loss: ', test_loss
log.write(buffer)
if test_loss < best_score:
best_score = test_loss
print 'best score: ', test_loss
significance_test(sess, target_lstm, likelihood_data_loader,
'significance/seqgan.txt')
rollout.update_params()
# generate for discriminator
print 'Start training discriminator'
for _ in range(5):
generate_samples(sess, generator, BATCH_SIZE, generated_num,
negative_file)
dis_x_train, dis_y_train = dis_data_loader.load_train_data(
positive_file, negative_file)
dis_batches = dis_data_loader.batch_iter(
zip(dis_x_train, dis_y_train), dis_batch_size, 3)
for batch in dis_batches:
try:
x_batch, y_batch = zip(*batch)
feed = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: dis_dropout_keep_prob
}
_, step = sess.run([dis_train_op, dis_global_step], feed)
except ValueError:
pass
log.close()
def main():
random.seed(SEED)
np.random.seed(SEED)
assert START_TOKEN == 0
gen_data_loader = Gen_Data_loader(BATCH_SIZE)
likelihood_data_loader = Likelihood_data_loader(BATCH_SIZE)
vocab_size = 5000
dis_data_loader = Dis_dataloader()
best_score = 1000
generator = get_trainable_model(vocab_size)
target_params = cPickle.load(open('save/target_params.pkl'))
target_lstm = TARGET_LSTM(vocab_size, 64, 32, 32, 20, 0, target_params)
with tf.variable_scope('discriminator'):
cnn = TextCNN(
sequence_length=20,
num_classes=2,
vocab_size=vocab_size,
embedding_size=dis_embedding_dim,
filter_sizes=dis_filter_sizes,
num_filters=dis_num_filters,
l2_reg_lambda=dis_l2_reg_lambda)
cnn_params = [param for param in tf.trainable_variables() if 'discriminator' in param.name]
# Define Discriminator Training procedure
dis_global_step = tf.Variable(0, name="global_step", trainable=False)
dis_optimizer = tf.train.AdamOptimizer(1e-4)
dis_grads_and_vars = dis_optimizer.compute_gradients(cnn.loss, cnn_params, aggregation_method=2)
dis_train_op = dis_optimizer.apply_gradients(dis_grads_and_vars, global_step=dis_global_step)
config = tf.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 0.5
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer)
generate_samples(sess, target_lstm, 64, 10000, positive_file)
gen_data_loader.create_batches(positive_file)
log = open('log/experiment-log.txt', 'w')
# pre-train generator
print 'Start pre-training...'
log.write('pre-training...\n')
for epoch in xrange(PRE_EPOCH_NUM):
print 'pre-train epoch:', epoch
loss = pre_train_epoch(sess, generator, gen_data_loader)
if epoch % 5 == 0:
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
print 'pre-train epoch ', epoch, 'test_loss ', test_loss
buffer = str(epoch) + ' ' + str(test_loss) + '\n'
log.write(buffer)
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = 'After pre-training:' + ' ' + str(test_loss) + '\n'
log.write(buffer)
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
significance_test(sess, target_lstm, likelihood_data_loader, 'significance/supervise.txt')
print 'Start training discriminator...'
for _ in range(dis_alter_epoch):
generate_samples(sess, generator, BATCH_SIZE, generated_num, negative_file)
# train discriminator
dis_x_train, dis_y_train = dis_data_loader.load_train_data(positive_file, negative_file)
dis_batches = dis_data_loader.batch_iter(
zip(dis_x_train, dis_y_train), dis_batch_size, dis_num_epochs
)
for batch in dis_batches:
try:
x_batch, y_batch = zip(*batch)
feed = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: dis_dropout_keep_prob
}
_, step = sess.run([dis_train_op, dis_global_step], feed)
except ValueError:
pass
rollout = ROLLOUT(generator, 0.8)
print '#########################################################################'
print 'Start Reinforcement Training Generator...'
log.write('Reinforcement Training...\n')
for total_batch in range(TOTAL_BATCH):
for it in range(TRAIN_ITER):
samples = generator.generate(sess)
rewards = rollout.get_reward(sess, samples, 16, cnn)
feed = {generator.x: samples, generator.rewards: rewards}
_, g_loss = sess.run([generator.g_updates, generator.g_loss], feed_dict=feed)
if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = str(total_batch) + ' ' + str(test_loss) + '\n'
print 'total_batch: ', total_batch, 'test_loss: ', test_loss
log.write(buffer)
if test_loss < best_score:
best_score = test_loss
print 'best score: ', test_loss
significance_test(sess, target_lstm, likelihood_data_loader, 'significance/seqgan.txt')
rollout.update_params()
# generate for discriminator
print 'Start training discriminator'
for _ in range(5):
generate_samples(sess, generator, BATCH_SIZE, generated_num, negative_file)
dis_x_train, dis_y_train = dis_data_loader.load_train_data(positive_file, negative_file)
dis_batches = dis_data_loader.batch_iter(zip(dis_x_train, dis_y_train), dis_batch_size, 3)
for batch in dis_batches:
try:
x_batch, y_batch = zip(*batch)
feed = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: dis_dropout_keep_prob
}
_, step = sess.run([dis_train_op, dis_global_step], feed)
except ValueError:
pass
log.close()
def main():
random.seed(SEED)
np.random.seed(SEED)
assert START_TOKEN == 0
gen_data_loader = Gen_Data_loader(BATCH_SIZE)
likelihood_data_loader = Likelihood_data_loader(BATCH_SIZE)
vocab_size = 5000
dis_data_loader = Dis_dataloader()
best_score = 1000
# initialize a LSTM object and use the LSTM object to initialize PoemGen object
generator = get_trainable_model(vocab_size)
# cPickle is a object serialization library
# the loaded picle object will be an array of numbers
# later, these params will be used to initalize the target LSTM
target_params = cPickle.load(open('save/target_params.pkl'))
# print target_params
time.sleep(1000)
# This seems like the generator model which used RNN
target_lstm = TARGET_LSTM(vocab_size, 64, 32, 32, 20, 0, target_params)
# This is the discriminator which uses CNN
with tf.variable_scope('discriminator'):
cnn = TextCNN(sequence_length=20,
num_classes=2,
vocab_size=vocab_size,
embedding_size=dis_embedding_dim,
filter_sizes=dis_filter_sizes,
num_filters=dis_num_filters,
l2_reg_lambda=dis_l2_reg_lambda)
cnn_params = [
param for param in tf.trainable_variables()
if 'discriminator' in param.name
]
# Define Discriminator Training procedure
dis_global_step = tf.Variable(0, name="global_step", trainable=False)
dis_optimizer = tf.train.AdamOptimizer(1e-4)
dis_grads_and_vars = dis_optimizer.compute_gradients(cnn.loss,
cnn_params,
aggregation_method=2)
dis_train_op = dis_optimizer.apply_gradients(dis_grads_and_vars,
global_step=dis_global_step)
config = tf.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 0.5
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.initialize_all_variables())
generate_samples(sess, target_lstm, 64, 10000, positive_file)
gen_data_loader.create_batches(positive_file)
log = open('log/experiment-log.txt', 'w')
# pre-train generator
# Initialize the generator with MLE estimators
print 'Start pre-training...'
log.write('pre-training...\n')
for epoch in xrange(PRE_EPOCH_NUM):
print 'pre-train epoch:', epoch
loss = pre_train_epoch(sess, generator, gen_data_loader)
if epoch % 5 == 0:
generate_samples(sess, generator, BATCH_SIZE, generated_num,
eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
print 'pre-train epoch ', epoch, 'test_loss ', test_loss
buffer = str(epoch) + ' ' + str(test_loss) + '\n'
log.write(buffer)
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = 'After pre-training:' + ' ' + str(test_loss) + '\n'
log.write(buffer)
generate_samples(sess, generator, BATCH_SIZE, generated_num, eval_file)
likelihood_data_loader.create_batches(eval_file)
significance_test(sess, target_lstm, likelihood_data_loader,
'significance/supervise.txt')
print 'Start training discriminator...'
for _ in range(dis_alter_epoch):
generate_samples(sess, generator, BATCH_SIZE, generated_num,
negative_file)
# train discriminator
dis_x_train, dis_y_train = dis_data_loader.load_train_data(
positive_file, negative_file)
dis_batches = dis_data_loader.batch_iter(zip(dis_x_train, dis_y_train),
dis_batch_size,
dis_num_epochs)
for batch in dis_batches:
try:
x_batch, y_batch = zip(*batch)
feed = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: dis_dropout_keep_prob
}
_, step = sess.run([dis_train_op, dis_global_step], feed)
except ValueError:
pass
rollout = ROLLOUT(generator, 0.8)
print '#########################################################################'
print 'Start Reinforcement Training Generator...'
log.write('Reinforcement Training...\n')
for total_batch in range(TOTAL_BATCH):
for it in range(TRAIN_ITER):
samples = generator.generate(sess)
rewards = rollout.get_reward(sess, samples, 16, cnn)
feed = {generator.x: samples, generator.rewards: rewards}
_, g_loss = sess.run([generator.g_updates, generator.g_loss],
feed_dict=feed)
if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
# The trainable model 'generator' is a RNN model from PoemGen
generate_samples(sess, generator, BATCH_SIZE, generated_num,
eval_file)
likelihood_data_loader.create_batches(eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = str(total_batch) + ' ' + str(test_loss) + '\n'
print 'total_batch: ', total_batch, 'test_loss: ', test_loss
log.write(buffer)
if test_loss < best_score:
best_score = test_loss
print 'best score: ', test_loss
significance_test(sess, target_lstm, likelihood_data_loader,
'significance/seqgan.txt')
# rollout policy???
rollout.update_params()
# generate for discriminator
print 'Start training discriminator'
for _ in range(5):
generate_samples(sess, generator, BATCH_SIZE, generated_num,
negative_file)
dis_x_train, dis_y_train = dis_data_loader.load_train_data(
positive_file, negative_file)
dis_batches = dis_data_loader.batch_iter(
zip(dis_x_train, dis_y_train), dis_batch_size, 3)
for batch in dis_batches:
try:
x_batch, y_batch = zip(*batch)
feed = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: dis_dropout_keep_prob
}
_, step = sess.run([dis_train_op, dis_global_step], feed)
except ValueError:
pass
log.close()