def create_seq2seq(session, mode):
if FLAGS.mode == 'TEST' or 'val' in FLAGS.mode:
FLAGS.schedule_sampling = False
else:
FLAGS.beam_search = False
print('FLAGS.beam_search: ',FLAGS.beam_search)
if FLAGS.beam_search:
print('FLAGS.beam_size: ',FLAGS.beam_size)
print('FLAGS.debug: ',bool(FLAGS.debug))
model = seq2seq_model.Seq2seq(mode)
if FLAGS.mode == 'val_rl':
ckpt = tf.train.get_checkpoint_state(FLAGS.model_rl_dir)
else:
ckpt = tf.train.get_checkpoint_state(FLAGS.model_pre_dir)
if FLAGS.load != '':
print("Reading model from %s, mode: %s" % (FLAGS.load, FLAGS.mode))
model.saver.restore(session, FLAGS.load)
elif ckpt:
print("Reading model from %s, mode: %s" % (ckpt.model_checkpoint_path, FLAGS.mode))
model.saver.restore(session, ckpt.model_checkpoint_path)
else:
print("Create model with fresh parameters, mode: %s" % FLAGS.mode)
session.run(tf.global_variables_initializer())
return model
def create_seq2seq(session, mode):
if mode == 'TEST':
FLAGS.schedule_sampling = False
else:
FLAGS.beam_search = False
print('FLAGS.beam_search: ',FLAGS.beam_search)
print('FLAGS.length_penalty: ',FLAGS.length_penalty)
print('FLAGS.length_penalty_factor: ',FLAGS.length_penalty_factor)
if FLAGS.beam_search:
print('FLAGS.beam_size: ',FLAGS.beam_size)
print('FLAGS.debug: ',bool(FLAGS.debug))
model = seq2seq_model.Seq2seq(src_vocab_size = FLAGS.src_vocab_size,
trg_vocab_size = FLAGS.trg_vocab_size,
buckets = buckets,
size = FLAGS.hidden_size,
num_layers = FLAGS.num_layers,
batch_size = FLAGS.batch_size,
mode = mode,
input_keep_prob = FLAGS.input_keep_prob,
output_keep_prob = FLAGS.output_keep_prob,
state_keep_prob = FLAGS.state_keep_prob,
beam_search = FLAGS.beam_search,
beam_size = FLAGS.beam_size,
schedule_sampling = FLAGS.schedule_sampling,
sampling_decay_rate = FLAGS.sampling_decay_rate,
sampling_global_step = FLAGS.sampling_global_step,
sampling_decay_steps = FLAGS.sampling_decay_steps,
pretrain_vec = FLAGS.pretrain_vec,
pretrain_trainable = FLAGS.pretrain_trainable,
length_penalty = FLAGS.length_penalty,
length_penalty_factor = FLAGS.length_penalty_factor
)
if len(FLAGS.bind) > 0:
ckpt = tf.train.get_checkpoint_state(FLAGS.bind)
elif mode != 'TEST':
if FLAGS.mode == "MLE":
ckpt = tf.train.get_checkpoint_state(FLAGS.model_dir)
elif FLAGS.mode == "RL":
ckpt = tf.train.get_checkpoint_state(FLAGS.model_rl_dir)
else:
if FLAGS.test_mode == "MLE":
ckpt = tf.train.get_checkpoint_state(FLAGS.model_dir)
elif FLAGS.test_mode == "RL":
ckpt = tf.train.get_checkpoint_state(FLAGS.model_rl_dir)
print("FLAGS.mode: ",FLAGS.mode)
if ckpt:
print("Reading model from %s, mode: %s" % (ckpt.model_checkpoint_path, mode))
model.saver.restore(session, ckpt.model_checkpoint_path)
else:
print("Create model with fresh parameters, mode: %s" % mode)
session.run(tf.global_variables_initializer())
return model
def create_seq2seq(session, mode):
if mode == 'TEST':
FLAGS.schedule_sampling = False
else:
FLAGS.beam_search = False
print('FLAGS.beam_search: ',FLAGS.beam_search)
if FLAGS.beam_search:
print('FLAGS.beam_size: ',FLAGS.beam_size)
print('FLAGS.debug: ',bool(FLAGS.debug))
model = seq2seq_model.Seq2seq(vocab_size = FLAGS.vocab_size,
buckets = buckets,
size = FLAGS.hidden_size,
num_layers = FLAGS.num_layers,
batch_size = FLAGS.batch_size,
mode = mode,
input_keep_prob = FLAGS.input_keep_prob,
output_keep_prob = FLAGS.output_keep_prob,
state_keep_prob = FLAGS.state_keep_prob,
beam_search = FLAGS.beam_search,
beam_size = FLAGS.beam_size,
schedule_sampling = FLAGS.schedule_sampling,
sampling_decay_rate = FLAGS.sampling_decay_rate,
sampling_global_step = FLAGS.sampling_global_step,
sampling_decay_steps = FLAGS.sampling_decay_steps
)
#if mode != 'TEST':
ckpt = tf.train.get_checkpoint_state(FLAGS.model_dir)
#else:
# ckpt = tf.train.get_checkpoint_state(FLAGS.model_rl_dir)
if ckpt:
print("Reading model from %s, mode: %s" % (ckpt.model_checkpoint_path, mode))
model.saver.restore(session, ckpt.model_checkpoint_path)
else:
print("Create model with fresh parameters, mode: %s" % mode)
session.run(tf.global_variables_initializer())
return model
def create_seq2seq(session, mode):
model = seq2seq_model.Seq2seq(vocab_size=FLAGS.vocab_size,
buckets=buckets,
size=FLAGS.hidden_size,
num_layers=FLAGS.num_layers,
batch_size=FLAGS.batch_size,
mode=mode)
#if mode != 'TEST':
ckpt = tf.train.get_checkpoint_state(FLAGS.model_dir)
#else:
# ckpt = tf.train.get_checkpoint_state(FLAGS.model_rl_dir)
if ckpt:
print("Reading model from %s, mode: %s" %
(ckpt.model_checkpoint_path, mode))
model.saver.restore(session, ckpt.model_checkpoint_path)
else:
print("Create model with fresh parameters, mode: %s" % mode)
session.run(tf.global_variables_initializer())
return model