def _init_subtokenizer(self, vocab_list):
temp_file = tempfile.NamedTemporaryFile(delete=False)
with tf.gfile.Open(temp_file.name, 'w') as w:
for subtoken in vocab_list:
w.write("'%s'" % subtoken)
w.write("\n")
return tokenizer.Subtokenizer(temp_file.name, reserved_tokens=[])
def translate_list(vocab, model_dir, params, contentList):
translation_results = []
subtokenizer = tokenizer.Subtokenizer(vocab)
estimator = tf.estimator.Estimator(model_fn=transformer_main.model_fn,
model_dir=model_dir,
params=params)
estimator_predictor = tf.contrib.predictor.from_estimator(
estimator,
export.build_tensor_serving_input_receiver_fn(shape=[None],
dtype=tf.int32,
batch_size=None))
for content in contentList:
try:
tokens = _encode_and_add_eos(content, subtokenizer)
predictions = estimator_predictor(
{"input": np.array([tokens], dtype=np.int32)})
translation = _trim_and_decode(predictions["outputs"][0],
subtokenizer)
translation_results.append(translation)
except:
print("error in translation")
return translation_results
def evaluate_and_log_bleu(model,
params,
bleu_source,
bleu_ref,
vocab_file,
distribution_strategy=None):
"""Calculate and record the BLEU score.
Args:
model: A Keras model, used to generate the translations.
params: A dictionary, containing the translation related parameters.
bleu_source: A file containing source sentences for translation.
bleu_ref: A file containing the reference for the translated sentences.
vocab_file: A file containing the vocabulary for translation.
distribution_strategy: A platform distribution strategy, used for TPU based
translation.
Returns:
uncased_score: A float, the case insensitive BLEU score.
cased_score: A float, the case sensitive BLEU score.
"""
subtokenizer = tokenizer.Subtokenizer(vocab_file)
uncased_score, cased_score = translate_and_compute_bleu(
model, params, subtokenizer, bleu_source, bleu_ref, distribution_strategy)
logging.info("Bleu score (uncased): %s", uncased_score)
logging.info("Bleu score (cased): %s", cased_score)
return uncased_score, cased_score
def main(unused_argv):
# changed to import transformer_main_hvd instead of transformer_main
from official.transformer import transformer_main_hvd
tf.logging.set_verbosity(tf.logging.INFO)
if FLAGS.text is None and FLAGS.file is None:
tf.logging.warn(
"Nothing to translate. Make sure to call this script using "
"flags --text or --file.")
return
subtokenizer = tokenizer.Subtokenizer(FLAGS.vocab_file)
# Set up estimator and params
params = transformer_main_hvd.PARAMS_MAP[FLAGS.param_set]
params["beam_size"] = _BEAM_SIZE
params["alpha"] = _ALPHA
params["extra_decode_length"] = _EXTRA_DECODE_LENGTH
params["batch_size"] = _DECODE_BATCH_SIZE
estimator = tf.estimator.Estimator(
model_fn=transformer_main_hvd.model_fn,
model_dir=FLAGS.model_dir,
params=params,
config=tf.estimator.RunConfig(session_config=tf.ConfigProto(
intra_op_parallelism_threads=FLAGS.intra_op,
inter_op_parallelism_threads=FLAGS.inter_op)))
# create translation directory
tf.gfile.MakeDirs(FLAGS.translations_dir)
if FLAGS.text is not None:
tf.logging.info("Translating text: %s" % FLAGS.text)
translate_text(estimator, subtokenizer, FLAGS.text)
if FLAGS.file is not None:
input_file = os.path.abspath(FLAGS.file)
tf.logging.info("Translating file: %s" % input_file)
if not tf.gfile.Exists(FLAGS.file):
raise ValueError("File does not exist: %s" % input_file)
""" output_file = None
if FLAGS.file_out is not None:
output_file = os.path.abspath(FLAGS.file_out)
tf.logging.info("File output specified: %s" % output_file) """
for model in checkpoint_yield.stepfiles_iterator(
FLAGS.model_dir,
wait_minutes=FLAGS.wait_minutes,
min_steps=FLAGS.min_steps):
checkpoint_path, checkpoint_file = os.path.split(model[0])
output_file = os.path.abspath(FLAGS.translations_dir + "/" +
checkpoint_file + "_" +
FLAGS.file_out)
tf.logging.info("Output file: %s" % output_file)
translate_file(model[0], estimator, subtokenizer, input_file,
output_file)
def evaluate_and_log_bleu(model, bleu_source, bleu_ref, vocab_file):
"""Calculate and record the BLEU score."""
subtokenizer = tokenizer.Subtokenizer(vocab_file)
uncased_score, cased_score = translate_and_compute_bleu(
model, subtokenizer, bleu_source, bleu_ref)
tf.compat.v1.logging.info("Bleu score (uncased): %s", uncased_score)
tf.compat.v1.logging.info("Bleu score (cased): %s", cased_score)
return uncased_score, cased_score
def evaluate(model,
params,
source_file,
target_file,
vocab_file,
distribution_strategy=None):
subtokenizer = tokenizer.Subtokenizer(vocab_file)
compute(model, params, subtokenizer, source_file, target_file,
distribution_strategy)
def main(unused_argv):
from official.transformer import transformer_main
tf.logging.set_verbosity(tf.logging.INFO)
if FLAGS.text is None and FLAGS.file is None:
tf.logging.warn(
"Nothing to translate. Make sure to call this script using "
"flags --text or --file.")
return
subtokenizer = tokenizer.Subtokenizer(FLAGS.vocab_file)
# Set up estimator and params
params = transformer_main.PARAMS_MAP[FLAGS.param_set]
params["beam_size"] = _BEAM_SIZE
params["alpha"] = _ALPHA
params["extra_decode_length"] = _EXTRA_DECODE_LENGTH
params["batch_size"] = _DECODE_BATCH_SIZE
params["concrete_coef"] = FLAGS.concrete_coef
params["concrete_heads"] = eval(FLAGS.concrete_heads)
params["alive_heads_enc_self"] = eval(FLAGS.alive_heads_enc_self)
params["alive_heads_dec_self"] = eval(FLAGS.alive_heads_dec_self)
params["alive_heads_enc_dec"] = eval(FLAGS.alive_heads_enc_dec)
print(" ******* Printing gate values ********")
print(params["alive_heads_enc_self"])
print(params["alive_heads_enc_dec"])
print(params["alive_heads_dec_self"])
print("concrete_coef : {}".format(params["concrete_coef"]))
print("concrete_heads : {}".format(params["concrete_heads"]))
estimator = tf.estimator.Estimator(model_fn=transformer_main.model_fn,
model_dir=FLAGS.model_dir,
params=params)
if FLAGS.text is not None:
tf.logging.info("Translating text: %s" % FLAGS.text)
translate_text(estimator, subtokenizer, FLAGS.text)
if FLAGS.file is not None:
input_file = os.path.abspath(FLAGS.file)
tf.logging.info("Translating file: %s" % input_file)
if not tf.gfile.Exists(FLAGS.file):
raise ValueError("File does not exist: %s" % input_file)
output_file = None
if FLAGS.file_out is not None:
output_file = os.path.abspath(FLAGS.file_out)
tf.logging.info("File output specified: %s" % output_file)
translate_file(estimator, subtokenizer, input_file, output_file)
def main(unused_argv):
from official.transformer import transformer_main
tf.logging.set_verbosity(tf.logging.INFO)
if FLAGS.text is None and FLAGS.file is None:
tf.logging.warn(
"Nothing to translate. Make sure to call this script using "
"flags --text or --file.")
return
subtokenizer = tokenizer.Subtokenizer(FLAGS.vocab_file)
# Set up estimator and params
params = transformer_main.PARAMS_MAP[FLAGS.param_set]
# debug
#print('debug: hidden_size %d' % params["hidden_size"])
#print('debug: filter_size %d' % params["filter_size"])
params["beam_size"] = _BEAM_SIZE
params["alpha"] = _ALPHA
params["extra_decode_length"] = _EXTRA_DECODE_LENGTH
params["batch_size"] = _DECODE_BATCH_SIZE
# TC: set vocab_size as the number of tokens in vocab_file
params["vocab_size"] = len(open(FLAGS.vocab_file).readlines())
print('TC: vocab_size %d' % params["vocab_size"])
estimator = tf.estimator.Estimator(model_fn=transformer_main.model_fn,
model_dir=FLAGS.model_dir,
params=params)
if FLAGS.text is not None:
tf.logging.info("Translating text: %s" % FLAGS.text)
translate_text(estimator, subtokenizer, FLAGS.text)
if FLAGS.file is not None:
input_file = os.path.abspath(FLAGS.file)
tf.logging.info("Translating file: %s" % input_file)
if not tf.gfile.Exists(FLAGS.file):
raise ValueError("File does not exist: %s" % input_file)
output_file = None
if FLAGS.file_out is not None:
output_file = os.path.abspath(FLAGS.file_out)
tf.logging.info("File output specified: %s" % output_file)
translate_file(estimator, subtokenizer, input_file, output_file)
def main(unused_argv):
from official.transformer import transformer_main
tf.logging.set_verbosity(tf.logging.INFO)
if FLAGS.text is None and FLAGS.file is None:
tf.logging.warn(
"Nothing to translate. Make sure to call this script using "
"flags --text or --file.")
return
subtokenizer = tokenizer.Subtokenizer(
os.path.join(FLAGS.data_dir, FLAGS.vocab_file))
if FLAGS.params == "base":
params = model_params.TransformerBaseParams
elif FLAGS.params == "big":
params = model_params.TransformerBigParams
else:
raise ValueError("Invalid parameter set defined: %s."
"Expected 'base' or 'big.'" % FLAGS.params)
# Set up estimator and params
params.beam_size = _BEAM_SIZE
params.alpha = _ALPHA
params.extra_decode_length = _EXTRA_DECODE_LENGTH
params.batch_size = _DECODE_BATCH_SIZE
estimator = tf.estimator.Estimator(model_fn=transformer_main.model_fn,
model_dir=FLAGS.model_dir,
params=params)
if FLAGS.text is not None:
tf.logging.info("Translating text: %s" % FLAGS.text)
translate_text(estimator, subtokenizer, FLAGS.text)
if FLAGS.file is not None:
input_file = os.path.abspath(FLAGS.file)
tf.logging.info("Translating file: %s" % input_file)
if not tf.gfile.Exists(FLAGS.file):
raise ValueError("File does not exist: %s" % input_file)
output_file = None
if FLAGS.file_out is not None:
output_file = os.path.abspath(FLAGS.file_out)
tf.logging.info("File output specified: %s" % output_file)
translate_file(estimator, subtokenizer, input_file, output_file)
def predict(self):
"""Predicts result from the model."""
params, flags_obj, is_train = self.params, self.flags_obj, False
with tf.name_scope("model"):
model = transformer.create_model(params, is_train)
self._load_weights_if_possible(model, flags_obj.init_weight_path)
model.summary()
subtokenizer = tokenizer.Subtokenizer(flags_obj.vocab_file)
ds = data_pipeline.eval_input_fn(params)
ds = ds.map(lambda x, y: x).take(_SINGLE_SAMPLE)
ret = model.predict(ds)
val_outputs, _ = ret
length = len(val_outputs)
for i in range(length):
translate.translate_from_input(val_outputs[i], subtokenizer)
def main(unused_argv):
from official.transformer import transformer_main_triblock as transformer_main
tf.logging.set_verbosity(tf.logging.INFO)
if FLAGS.text is None and FLAGS.file is None:
tf.logging.warn(
"Nothing to translate. Make sure to call this script using "
"flags --text or --file.")
return
underscored_ids = FLAGS.underscored_ids.split(",")
underscored_ids = [int(idx) for idx in underscored_ids]
subtokenizer = tokenizer.Subtokenizer(FLAGS.vocab_file)
#tf.logging.info(subtokenizer.subtoken_list[:] )
# Set up estimator and params
params = transformer_main.PARAMS_MAP[FLAGS.param_set]
params["underscored_ids"] = underscored_ids
params["vocab_file"] = FLAGS.vocab_file
params["beam_size"] = _BEAM_SIZE
params["alpha"] = _ALPHA
params["extra_decode_length"] = _EXTRA_DECODE_LENGTH
params["batch_size"] = _DECODE_BATCH_SIZE
estimator = tf.estimator.Estimator(model_fn=transformer_main.model_fn,
model_dir=FLAGS.model_dir,
params=params)
if FLAGS.text is not None:
tf.logging.info("Translating text: %s" % FLAGS.text)
translate_text(estimator, subtokenizer, FLAGS.text)
if FLAGS.file is not None:
input_file = os.path.abspath(FLAGS.file)
tf.logging.info("Translating file: %s" % input_file)
if not tf.gfile.Exists(FLAGS.file):
raise ValueError("File does not exist: %s" % input_file)
output_file = None
if FLAGS.file_out is not None:
output_file = os.path.abspath(FLAGS.file_out)
tf.logging.info("File output specified: %s" % output_file)
translate_file(estimator, subtokenizer, input_file, output_file)
def evaluate_and_log_bleu(estimator, bleu_writer, bleu_source, bleu_ref):
"""Calculate and record the BLEU score."""
subtokenizer = tokenizer.Subtokenizer(
os.path.join(FLAGS.data_dir, FLAGS.vocab_file))
uncased_score, cased_score = translate_and_compute_bleu(
estimator, subtokenizer, bleu_source, bleu_ref)
print("Bleu score (uncased):", uncased_score)
print("Bleu score (cased):", cased_score)
summary = tf.Summary(value=[
tf.Summary.Value(tag="bleu/uncased", simple_value=uncased_score),
tf.Summary.Value(tag="bleu/cased", simple_value=cased_score),
])
bleu_writer.add_summary(summary, get_global_step(estimator))
bleu_writer.flush()
return uncased_score, cased_score
def predict(self):
"""Predicts result from the model."""
self.params['train'] = False
params = self.params
flags_obj = self.flags_obj
with tf.name_scope("model"):
model = transformer.create_model(params, is_train=False)
self._load_weights_if_possible(
model, tf.train.latest_checkpoint(self.flags_obj.model_dir))
model.summary()
subtokenizer = tokenizer.Subtokenizer(flags_obj.vocab_file)
print(params)
ds = data_pipeline.eval_input_fn(params)
ds = ds.map(lambda x, y: x).take(_SINGLE_SAMPLE)
import time
start = time.time()
ret = model.predict(ds)
val_outputs, _ = ret
length = len(val_outputs)
for i in range(length):
translate.translate_from_input(val_outputs[i], subtokenizer)
print('\n\n\n', time.time() - start)
def predict(self, encoder_outputs, encoder_decoder_attention_bias):
"""Return predicted sequence."""
batch_size = tf.shape(encoder_outputs)[0]
input_length = tf.shape(encoder_outputs)[1]
max_decode_length = self.params["max_output_length"]
symbols_to_logits_fn = self._get_symbols_to_logits_fn(
max_decode_length)
# Create initial set of IDs that will be passed into symbols_to_logits_fn.
initial_ids = tf.zeros([batch_size], dtype=tf.int32)
# Create cache storing decoder attention values for each layer.
cache = {
"layer_%d" % layer: {
"k": tf.zeros([batch_size, 0, self.params["hidden_size"]]),
"v": tf.zeros([batch_size, 0, self.params["hidden_size"]]),
}
for layer in range(self.params["num_hidden_layers"])
}
# Add encoder output and attention bias to the cache.
cache["encoder_outputs"] = encoder_outputs
cache[
"encoder_decoder_attention_bias"] = encoder_decoder_attention_bias
####domyoung 2019.12.9#####
nontrigrams = nontrigram_generator(max_decode_length,
self.params["underscored_ids"])
nontrigrams = tf.constant(nontrigrams, dtype=tf.int32)
tile_dims = [1] * nontrigrams.shape.ndims
tile_dims[-1] = batch_size * self.params["beam_size"]
nontrigrams = tf.tile(nontrigrams, tile_dims)
nontrigrams = tf.reshape(nontrigrams, [-1, max_decode_length])
subtokenizer = tokenizer.Subtokenizer(self.params["vocab_file"])
key = tf.range(self.params["vocab_size"], dtype=tf.int32)
#replace the first token '<pad>_' into '<pad>'
subtoken_list = subtokenizer.subtoken_list[:]
value = tf.constant(subtoken_list, dtype=tf.string)
default_value = tf.constant("", dtype=tf.string)
hashTable = tf.contrib.lookup.HashTable(
tf.contrib.lookup.KeyValueTensorInitializer(key, value),
default_value)
# Use beam search to find the top beam_size sequences and scores.
tf.logging.info(key)
tf.logging.info(value)
tf.logging.info(subtoken_list)
decoded_ids, scores = beam_search.sequence_beam_search(
symbols_to_logits_fn=symbols_to_logits_fn,
initial_ids=initial_ids,
initial_cache=cache,
vocab_size=self.params["vocab_size"],
hashTable=hashTable,
nontrigrams=nontrigrams,
use_trigram=True,
beam_size=self.params["beam_size"],
batch_size=self.params["batch_size"],
alpha=self.params["alpha"],
max_decode_length=max_decode_length,
eos_id=EOS_ID)
#########################
# Get the top sequence for each batch element
top_decoded_ids = decoded_ids[:, 0, 1:]
top_scores = scores[:, 0]
return {"outputs": top_decoded_ids, "scores": top_scores}
