def score_model(source_file, target_file, scorer_settings, options):
scores = []
for option in options:
g = tf.Graph()
with g.as_default():
with tf.Session() as sess:
model, saver = nmt.create_model(option, sess)
text_iterator = TextIterator(
source=source_file.name,
target=target_file.name,
source_dicts=option.source_dicts,
target_dict=option.target_dict,
batch_size=scorer_settings.b,
maxlen=float('inf'),
source_vocab_sizes=option.source_vocab_sizes,
target_vocab_size=option.target_vocab_size,
use_factor=(option.factors > 1),
sort_by_length=False)
losses = nmt.calc_loss_per_sentence(
option,
sess,
text_iterator,
model,
normalization_alpha=scorer_settings.normalization_alpha)
scores.append(losses)
return scores
def score_model(source_file, target_file, scorer_settings, options):
scores = []
for option in options:
with tf.Session() as sess:
model, saver = create_model(option, sess)
valid_text_iterator = TextIterator(
source=source_file.name,
target=target_file.name,
source_dicts=option.source_dicts,
target_dict=option.target_dict,
batch_size=scorer_settings.b,
maxlen=float('inf'),
source_vocab_sizes=option.source_vocab_sizes,
target_vocab_size=option.target_vocab_size,
use_factor=(option.factors > 1),
sort_by_length=False)
score = validate(
option,
sess,
valid_text_iterator,
model,
normalization_alpha=scorer_settings.normalization_alpha)
scores.append(score)
return scores
def theano_to_tensorflow_model(in_path, out_path):
saved_model = np.load(in_path)
config = theano_to_tensorflow_config(in_path)
th2tf = construct_parameter_map(config)
with tf.Session() as sess:
model, saver = nmt.create_model(config, sess)
seen = set()
assign_ops = []
for key in saved_model.keys():
if key in th2tf:
tf_name = th2tf[key]
if tf_name is not None:
assert tf_name not in seen
seen.add(tf_name)
tf_var = tf.get_default_graph().get_tensor_by_name(tf_name)
if (sess.run(tf.shape(tf_var)) !=
saved_model[key].shape).any():
print "mismatch for", tf_name, key, saved_model[
key].shape, sess.run(tf.shape(tf_var))
assign_ops.append(tf.assign(tf_var, saved_model[key]))
else:
print "Not saving", key, "because no TF equivalent"
sess.run(assign_ops)
saver.save(sess, save_path=out_path)
print "The following TF variables were not assigned (excluding Adam vars):"
print "You should see only 'beta1_power', 'beta2_power' and 'time' variable listed"
for tf_var in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES):
if tf_var.name not in seen and 'Adam' not in tf_var.name:
print tf_var.name
def theano_to_tensorflow_model(in_path, out_path):
import numpy as np
import tensorflow as tf
from nmt import create_model
saved_model = np.load(in_path)
# Create fake config
src_size, first_src_emb_size = saved_model['Wemb'].shape
dim_per_factor = [first_src_emb_size]
# check for additional source embedding layers (one per factor)
i = 1
while True:
th_name = 'Wemb{0}'.format(i)
if th_name not in saved_model:
break
dim_per_factor.append(saved_model[th_name].shape[1])
th2tf[th_name] = 'encoder/embedding/embeddings_{0}:0'.format(i)
i += 1
src_emb_size = sum(dim_per_factor)
trg_size, trg_emb_size = saved_model['Wemb_dec'].shape
state_size = saved_model['decoder_U'].shape[0]
if 'encoder_Wx_lns' in saved_model:
layer_norm = True
else:
layer_norm = False
if 'ff_logit_W' in saved_model:
tie_decoder_embeddings = False
else:
tie_decoder_embeddings = True
assert trg_emb_size == src_emb_size, 'src_emb_size ({}) and trg_emb_size ({}) do not equal, this is unsupported in TF'.format(
src_emb_size, trg_emb_size)
fake_config = FakeConfig(state_size, dim_per_factor, src_size, trg_size,
layer_norm, tie_decoder_embeddings)
with tf.Session() as sess:
model, saver = create_model(fake_config, sess)
seen = set()
assign_ops = []
for key in saved_model.keys():
tf_name = th2tf[key]
if tf_name is not None:
assert tf_name not in seen
seen.add(tf_name)
tf_var = tf.get_default_graph().get_tensor_by_name(tf_name)
if (sess.run(tf.shape(tf_var)) !=
saved_model[key].shape).any():
print "mismatch for", tf_name, key, saved_model[
key].shape, sess.run(tf.shape(tf_var))
assign_ops.append(tf.assign(tf_var, saved_model[key]))
else:
print "Not saving", key, "because no TF equivalent"
sess.run(assign_ops)
saver.save(sess, save_path=out_path)
print "The following TF variables were not assigned (excluding Adam vars):"
print "You should see only 'beta1_power', 'beta2_power' and 'time' variable listed"
for tf_var in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES):
if tf_var.name not in seen and 'Adam' not in tf_var.name:
print tf_var.name
def _load_models(self, process_id, sess):
"""
Loads models and returns them
"""
logging.debug("Process '%s' - Loading models\n" % (process_id))
import tensorflow as tf
models = []
for i, options in enumerate(self._options):
with tf.variable_scope("model%d" % i) as scope:
model, saver = create_model(options, sess, ensemble_scope=scope)
models.append(model)
logging.info("NOTE: Length of translations is capped to {}".format(self._options[0].translation_maxlen))
return models