import argparse
parser = argparse.ArgumentParser()
# for word segmentation
parser.add_argument("--source_input", default='corpus/source_input', dest="src_inp", help="source input file for word segmentation")
parser.add_argument("--target_input", default='corpus/target_input', dest="trg_inp", help="target input file for word segmentation")
parser.add_argument("--source_mode", default='word', choices=['char', 'word'], dest="src_mode", help="char or word")
parser.add_argument("--target_mode", default='char', choices=['char', 'word'], dest="trg_mode", help="char or word")
# for dataset splitting (train & val)
parser.add_argument('--source', default=FLAGS.source_data, dest='src', help='source file')
parser.add_argument('--target', default=FLAGS.target_data, dest='trg', help='target file')
args = parser.parse_args()
#step 0. apply word segmentation
## source
data_utils.word_seg(args.src_inp,args.src,args.src_mode.lower())
## target
data_utils.word_seg(args.trg_inp,args.trg,args.trg_mode.lower())
#step 1. get mapping of whole data
data_utils.prepare_whole_data(FLAGS.source_data, FLAGS.target_data, FLAGS.src_vocab_size, FLAGS.trg_vocab_size, skip_to_token=True)
#step 2. split data into train & val
data_utils.split_train_val(args.src,args.trg)
#step 3. generate tokens of train & val
data_utils.prepare_whole_data(FLAGS.source_data, FLAGS.target_data, FLAGS.src_vocab_size, FLAGS.trg_vocab_size, mode=word_seg_strategy)
#step 4. pretrain fasttext
data_utils.train_fasttext(fasttext_model,source_mapping,fasttext_hkl)
def train_MLE():
data_utils.prepare_whole_data(FLAGS.data, FLAGS.data_test, FLAGS.source_data, FLAGS.target_data, FLAGS.src_vocab_size, FLAGS.trg_vocab_size)
_ , trg_vocab_list = data_utils.read_map(FLAGS.target_data + '.' + str(FLAGS.trg_vocab_size) + '.mapping')
d_train = data_utils.read_data(FLAGS.source_data + '_train.token',FLAGS.target_data + '_train.token',buckets)
d_valid = data_utils.read_data(FLAGS.source_data + '_val.token',FLAGS.target_data + '_val.token',buckets)
print('Total document size of training data: %s' % sum(len(l) for l in d_train))
print('Total document size of validation data: %s' % sum(len(l) for l in d_valid))
train_bucket_sizes = [len(d_train[b]) for b in range(len(d_train))]
train_total_size = float(sum(train_bucket_sizes))
train_buckets_scale = [sum(train_bucket_sizes[:i + 1]) / train_total_size
for i in range(len(train_bucket_sizes))]
print('train_bucket_sizes: ',train_bucket_sizes)
print('train_total_size: ',train_total_size)
print('train_buckets_scale: ',train_buckets_scale)
valid_bucket_sizes = [len(d_valid[b]) for b in range(len(d_valid))]
valid_total_size = float(sum(valid_bucket_sizes))
valid_buckets_scale = [sum(valid_bucket_sizes[:i + 1]) / valid_total_size
for i in range(len(valid_bucket_sizes))]
print('valid_bucket_sizes: ',valid_bucket_sizes)
print('valid_total_size: ',valid_total_size)
print('valid_buckets_scale: ',valid_buckets_scale)
with tf.Session() as sess:
model = create_seq2seq(sess, 'MLE')
if FLAGS.reset_sampling_prob:
with tf.variable_scope('sampling_prob',reuse=tf.AUTO_REUSE):
sess.run(tf.assign(model.sampling_probability,reset_prob))
if FLAGS.schedule_sampling:
print('model.sampling_probability: ',model.sampling_probability_clip)
#sess.run(tf.assign(model.sampling_probability,1.0))
step = 0
loss = 0
loss_list = []
if FLAGS.schedule_sampling:
print('sampling_decay_steps: ',FLAGS.sampling_decay_steps)
print('sampling_probability: ',sess.run(model.sampling_probability_clip))
print('-----')
while step < FLAGS.max_step:
step += 1
random_number = np.random.random_sample()
# buckets_scale accumulated percentage
bucket_id = min([i for i in range(len(train_buckets_scale))
if train_buckets_scale[i] > random_number])
encoder_input, decoder_input, weight, en_s, de_s = model.get_batch(d_train, bucket_id, sen=True)
#print('batch_size: ',model.batch_size) ==> 64
#print('batch_size: ',len(encoder_input[0])) ==> 64
#print('batch_size: ',len(encoder_input)) ==> 15,50,...
#print('batch_size: ',len(decoder_input)) ==> 15,50,...
#print('batch_size: ',len(weight)) ==> 15,50,...
output, loss_train, _ = model.run(sess, encoder_input, decoder_input, weight, bucket_id)
loss += loss_train / FLAGS.check_step
#if step!=0 and step % FLAGS.sampling_decay_steps == 0:
# sess.run(model.sampling_probability_decay)
# print('sampling_probability: ',sess.run(model.sampling_probability))
if step % FLAGS.print_step == 0:
print('Input :')
print(en_s[0].strip())
print('Output:')
print(_output(output[0], trg_vocab_list))
print('\n{} steps trained ...\n\n'.format(step))
if step % FLAGS.check_step == 0:
print('\nStep %s, Training perplexity: %s, Learning rate: %s' % (step, math.exp(loss),
sess.run(model.learning_rate)))
for i in range(len(d_train)):
encoder_input, decoder_input, weight = model.get_batch(d_valid, i)
_, loss_valid = model.run(sess, encoder_input, decoder_input, weight, i, forward_only = True)
print(' Validation perplexity in bucket %s: %s' % (i, math.exp(loss_valid)))
if len(loss_list) > 2 and loss > max(loss_list[-3:]):
sess.run(model.learning_rate_decay)
else:
if step!=0:
if FLAGS.schedule_sampling:
sess.run(model.sampling_probability_decay)
print('sampling_probability: ',sess.run(model.sampling_probability_clip))
loss_list.append(loss)
loss = 0
checkpoint_path = os.path.join(FLAGS.model_pre_dir, "MLE.ckpt")
model.saver.save(sess, checkpoint_path, global_step = step)
print('Saving model at step %s\n' % step)
if step == FLAGS.sampling_global_step: break
def train_MLE():
data_utils.prepare_whole_data(FLAGS.source_data_dir, FLAGS.target_data_dir, FLAGS.vocab_size)
# read dataset and split to training set and validation set
d = data_utils.read_data(FLAGS.source_data_dir + '.token', FLAGS.target_data_dir + '.token', buckets)
np.random.seed(SEED)
np.random.shuffle(d)
print('Total document size: %s' % sum(len(l) for l in d))
print('len(d): ', len(d))
d_train = [[] for _ in range(len(d))]
d_valid = [[] for _ in range(len(d))]
for i in range(len(d)):
d_train[i] = d[i][:int(0.9 * len(d[i]))]
d_valid[i] = d[i][int(-0.1 * len(d[i])):]
train_bucket_sizes = [len(d_train[b]) for b in range(len(d_train))]
train_total_size = float(sum(train_bucket_sizes))
train_buckets_scale = [sum(train_bucket_sizes[:i + 1]) / train_total_size
for i in range(len(train_bucket_sizes))]
print('train_bucket_sizes: ',train_bucket_sizes)
print('train_total_size: ',train_total_size)
print('train_buckets_scale: ',train_buckets_scale)
valid_bucket_sizes = [len(d_valid[b]) for b in range(len(d_valid))]
valid_total_size = float(sum(valid_bucket_sizes))
valid_buckets_scale = [sum(valid_bucket_sizes[:i + 1]) / valid_total_size
for i in range(len(valid_bucket_sizes))]
print('valid_bucket_sizes: ',valid_bucket_sizes)
print('valid_total_size: ',valid_total_size)
print('valid_buckets_scale: ',valid_buckets_scale)
with tf.Session() as sess:
model = create_seq2seq(sess, 'MLE')
step = 0
loss = 0
loss_list = []
print('sampling_decay_steps: ',FLAGS.sampling_decay_steps)
print('sampling_probability: ',sess.run(model.sampling_probability))
print('-----')
while(True):
step += 1
random_number = np.random.random_sample()
# buckets_scale 是累加百分比
bucket_id = min([i for i in range(len(train_buckets_scale))
if train_buckets_scale[i] > random_number])
encoder_input, decoder_input, weight = model.get_batch(d_train, bucket_id)
#print('batch_size: ',model.batch_size) ==> 64
#print('batch_size: ',len(encoder_input[0])) ==> 64
#print('batch_size: ',len(encoder_input)) ==> 15,50,...
#print('batch_size: ',len(decoder_input)) ==> 15,50,...
#print('batch_size: ',len(weight)) ==> 15,50,...
loss_train, _ = model.run(sess, encoder_input, decoder_input, weight, bucket_id)
loss += loss_train / FLAGS.check_step
#print(model.token2word(sen)[0])
if step!=0 and step % FLAGS.sampling_decay_steps == 0:
sess.run(model.sampling_probability_decay)
print('sampling_probability: ',sess.run(model.sampling_probability))
if_feed_prev = bernoulli_sampling(model.sampling_probability)
if_feed_prev = sess.run(if_feed_prev)
print('if_feed_prev: ',not if_feed_prev)
if step % FLAGS.check_step == 0:
print('Step %s, Training perplexity: %s, Learning rate: %s' % (step, math.exp(loss),
sess.run(model.learning_rate)))
for i in range(len(d)):
encoder_input, decoder_input, weight = model.get_batch(d_valid, i)
loss_valid, _ = model.run(sess, encoder_input, decoder_input, weight, i, forward_only = True)
print(' Validation perplexity in bucket %s: %s' % (i, math.exp(loss_valid)))
if len(loss_list) > 2 and loss > max(loss_list[-3:]):
sess.run(model.learning_rate_decay)
loss_list.append(loss)
loss = 0
checkpoint_path = os.path.join(FLAGS.model_dir, "MLE.ckpt")
model.saver.save(sess, checkpoint_path, global_step = step)
print('Saving model at step %s' % step)
def train_RL():
data_utils.prepare_whole_data(FLAGS.data, FLAGS.data_test, FLAGS.source_data, FLAGS.target_data, FLAGS.src_vocab_size, FLAGS.trg_vocab_size)
d_train = data_utils.read_data(FLAGS.source_data + '_train.token',FLAGS.target_data + '_train.token',buckets)
#print(d_train[0][0])
g1 = tf.Graph()
g2 = tf.Graph()
g3 = tf.Graph()
sess1 = tf.Session(graph = g1)
sess2 = tf.Session(graph = g2)
sess3 = tf.Session(graph = g3)
# model is for training seq2seq with Reinforcement Learning
with g1.as_default():
model = create_seq2seq(sess1, 'RL')
# we set sample size = ?
model.batch_size = 5
# model_LM is for a reward function (language model)
with g2.as_default():
model_LM = create_seq2seq(sess2, 'MLE')
model_LM.beam_search = False
# calculate probibility of only one sentence
model_LM.batch_size = 1
def LM(encoder_input, decoder_input, weight, bucket_id):
return model_LM.run(sess2, encoder_input, decoder_input, weight, bucket_id, forward_only = True)[0]
# new reward function: sentiment score
with g3.as_default():
model_SA = main.create_model(sess3)
model_SA.batch_size = 1
def SA(sentence, encoder_length):
sentence = ' '.join(sentence)
token_ids = utils.convert_to_token(sentence, model_SA.vocab_map)
encoder_input, encoder_length, _, _ = model_SA.get_batch([(0, token_ids, sentence)])
return model_SA.step(sess3, encoder_input, encoder_length)[0][0]
train_bucket_sizes = [len(d_train[b]) for b in range(len(d_train))]
train_total_size = float(sum(train_bucket_sizes))
train_buckets_scale = [sum(train_bucket_sizes[:i + 1]) / train_total_size
for i in range(len(train_bucket_sizes))]
# make RL object read vocab mapping dict, list
model.RL_readmap(FLAGS.source_data + '.' + str(FLAGS.src_vocab_size) + '.mapping', FLAGS.target_data + '.' + str(FLAGS.trg_vocab_size) + '.mapping')
step = 0
while step < FLAGS.max_step:
step += 1
random_number = np.random.random_sample()
bucket_id = min([i for i in range(len(train_buckets_scale))
if train_buckets_scale[i] > random_number])
# the same encoder_input for sampling batch_size times
#encoder_input, decoder_input, weight = model.get_batch(d, bucket_id, rand = False)
encoder_input, decoder_input, weight, en_s, de_s = model.get_batch(d_train, bucket_id, sen=True)
output, loss, _ = model.run(sess1, encoder_input, decoder_input, weight, bucket_id, X = LM, Y = SA)
# debug
#encoder_input = np.reshape(np.transpose(encoder_input, (1, 0, 2)), (-1, FLAGS.vocab_size))
#encoder_input = np.split(encoder_input, FLAGS.max_length)
#print(model.token2word(encoder_input)[0])
#print(model.token2word(sen)[0])
if step % FLAGS.print_step == 0:
print('Input :')
print(en_s[0].strip())
print('Output:')
print(_output(output[0], model.trg_vocab_list))
print('\n{} steps trained ...'.format(step))
if step % FLAGS.check_step == 0:
print('Loss at step %s: %s' % (step, loss))
checkpoint_path = os.path.join(FLAGS.model_rl_dir, "RL.ckpt")
model.saver.save(sess1, checkpoint_path, global_step = step)
print('Saving model at step %s' % step)
def train_MLE():
data_utils.prepare_whole_data(FLAGS.source_data_dir, FLAGS.target_data_dir,
FLAGS.vocab_size)
# read dataset and split to training set and validation set
d = data_utils.read_data(FLAGS.source_data_dir + '.token',
FLAGS.target_data_dir + '.token', buckets)
print('Total document size: %s' % sum(len(l) for l in d))
d_train = [[] for _ in range(len(d))]
d_valid = [[] for _ in range(len(d))]
for i in range(len(d)):
d_train[i] = d[i][:int(0.9 * len(d[i]))]
d_valid[i] = d[i][int(-0.1 * len(d[i])):]
train_bucket_sizes = [len(d[b]) for b in range(len(d))]
train_total_size = float(sum(train_bucket_sizes))
train_buckets_scale = [
sum(train_bucket_sizes[:i + 1]) / train_total_size
for i in range(len(train_bucket_sizes))
]
sess = tf.Session()
model = create_seq2seq(sess, 'MLE')
step = 0
loss = 0
loss_list = []
while (True):
step += 1
random_number = np.random.random_sample()
bucket_id = min([
i for i in range(len(train_buckets_scale))
if train_buckets_scale[i] > random_number
])
encoder_input, decoder_input, weight = model.get_batch(
d_train, bucket_id)
loss_train, _ = model.run(sess, encoder_input, decoder_input, weight,
bucket_id)
loss += loss_train / FLAGS.check_step
#print(model.token2word(sen)[0])
if step % FLAGS.check_step == 0:
print('Step %s, Training perplexity: %s, Learning rate: %s' %
(step, math.exp(loss), sess.run(model.learning_rate)))
for i in range(len(d)):
encoder_input, decoder_input, weight = model.get_batch(
d_valid, i)
loss_valid, _ = model.run(sess,
encoder_input,
decoder_input,
weight,
i,
forward_only=True)
print(' Validation perplexity in bucket %s: %s' %
(i, math.exp(loss_valid)))
if len(loss_list) > 2 and loss > max(loss_list[-3:]):
sess.run(model.learning_rate_decay)
loss_list.append(loss)
loss = 0
checkpoint_path = os.path.join(FLAGS.model_dir, "MLE.ckpt")
model.saver.save(sess, checkpoint_path, global_step=step)
print('Saving model at step %s' % step)