def main():
utils.heading('SETUP')
config = configure.Config(mode=FLAGS.mode, model_name=FLAGS.model_name)
config.write()
with tf.Graph().as_default() as graph:
model_trainer = trainer.Trainer(config)
summary_writer = tf.summary.FileWriter(config.summaries_dir)
checkpoints_saver = tf.train.Saver(max_to_keep=1)
best_model_saver = tf.train.Saver(max_to_keep=1)
init_op = tf.global_variables_initializer()
graph.finalize()
with tf.Session() as sess:
sess.run(init_op)
progress = training_progress.TrainingProgress(
config, sess, checkpoints_saver, best_model_saver,
config.mode == 'train')
utils.log()
if config.mode == 'train':
utils.heading('START TRAINING ({:})'.format(config.model_name))
model_trainer.train(sess, progress, summary_writer)
elif config.mode == 'eval':
utils.heading('RUN EVALUATION ({:})'.format(config.model_name))
progress.best_model_saver.restore(
sess, tf.train.latest_checkpoint(config.checkpoints_dir))
model_trainer.evaluate_all_tasks(sess, summary_writer, None)
else:
raise ValueError('Mode must be "train" or "eval"')
def main(data_dir='/content/data'):
random.seed(0)
utils.log("BUILDING WORD VOCABULARY/EMBEDDINGS")
for pretrained in ['glove.6B.50d.txt']:
config = configure.Config(data_dir=data_dir,
for_preprocessing=True,
pretrained_embeddings=pretrained,
word_embedding_size=50)
embeddings.PretrainedEmbeddingLoader(config).build()
utils.log("CONSTRUCTING DEV SETS")
for task_name in ["chunk"]:
# chunking does not come with a provided dev split, so create one by
# selecting a random subset of the data
config = configure.Config(data_dir=data_dir,
for_preprocessing=True)
task_data_dir = os.path.join(config.raw_data_topdir, task_name) + '/'
train_sentences = word_level_data.TaggedDataLoader(
config, task_name, False).get_labeled_sentences("train")
random.shuffle(train_sentences)
write_sentences(task_data_dir + 'train_subset.txt', train_sentences[1500:])
write_sentences(task_data_dir + 'dev.txt', train_sentences[:1500])
utils.log("WRITING LABEL MAPPINGS")
for task_name in ["chunk"]:
for i, label_encoding in enumerate(["BIOES"]):
config = configure.Config(data_dir=data_dir,
for_preprocessing=True,
label_encoding=label_encoding)
token_level = task_name in ["ccg", "pos", "depparse"]
loader = word_level_data.TaggedDataLoader(config, task_name, token_level)
if token_level:
if i != 0:
continue
utils.log("WRITING LABEL MAPPING FOR", task_name.upper())
else:
utils.log(" Writing label mapping for", task_name.upper(),
label_encoding)
utils.log(" ", len(loader.label_mapping), "classes")
utils.write_cpickle(loader.label_mapping,
loader.label_mapping_path)
def main(data_dir='/content/data'):
random.seed(0)
utils.log("BUILDING WORD VOCABULARY/EMBEDDINGS")
for pretrained in ['glove.6B.100d.txt']:
config = configure.Config(data_dir=data_dir,
for_preprocessing=True,
pretrained_embeddings=pretrained,
word_embedding_size=100)
embeddings.PretrainedEmbeddingLoader(config).build()
utils.log("WRITING LABEL MAPPINGS")
for task_name in ["senclass"]:
config = configure.Config(data_dir=data_dir,
for_preprocessing=True)
loader = sentence_level_data.SentenceClassificationDataLoader(config, task_name)
utils.log("WRITING LABEL MAPPING FOR", task_name.upper())
utils.log(" ", len(loader.label_mapping), "classes")
utils.write_cpickle(loader.label_mapping,
loader.label_mapping_path)
def main():
utils.heading('SETUP')
config = configure.Config(mode=FLAGS.mode, model_name=FLAGS.model_name)
config.write()
if config.mode == 'encode':
word_vocab = embeddings.get_word_vocab(config)
sentence = "Squirrels , for example , would show up , look for the peanut , go away .".split()
sentence = ([word_vocab[embeddings.normalize_word(w)] for w in sentence])
print(sentence)
return
if config.mode == 'decode':
word_vocab_reversed = embeddings.get_word_vocab_reversed(config)
sentence = "25709 33 42 879 33 86 304 92 33 676 42 32 13406 33 273 445 34".split()
sentence = ([word_vocab_reversed[int(w)] for w in sentence])
print(sentence)
return
if config.mode == 'encode-vi':
word_vocab_vi = embeddings.get_word_vocab_vi(config)
print(len(word_vocab_vi))
sentence = "Mỗi_một khoa_học_gia đều thuộc một nhóm nghiên_cứu , và mỗi nhóm đều nghiên_cứu rất nhiều đề_tài đa_dạng .".split()
sentence = ([word_vocab_vi[embeddings.normalize_word(w)] for w in sentence])
print(sentence)
return
if config.mode == 'decode-vi':
word_vocab_reversed_vi = embeddings.get_word_vocab_reversed_vi(config)
sentence = "8976 32085 129 178 17 261 381 5 7 195 261 129 381 60 37 2474 1903 6".split()
sentence = ([word_vocab_reversed_vi[int(w)] for w in sentence])
print(sentence)
return
if config.mode == 'embed':
word_embeddings = embeddings.get_word_embeddings(config)
word = 50
embed = word_embeddings[word]
print(' '.join(str(x) for x in embed))
return
if config.mode == 'embed-vi':
word_embeddings_vi = embeddings.get_word_embeddings_vi(config)
word = 50
embed = word_embeddings_vi[word]
print(' '.join(str(x) for x in embed))
return
with tf.Graph().as_default() as graph:
model_trainer = trainer.Trainer(config)
summary_writer = tf.summary.FileWriter(config.summaries_dir)
checkpoints_saver = tf.train.Saver(max_to_keep=1)
best_model_saver = tf.train.Saver(max_to_keep=1)
init_op = tf.global_variables_initializer()
graph.finalize()
with tf.Session() as sess:
sess.run(init_op)
progress = training_progress.TrainingProgress(
config, sess, checkpoints_saver, best_model_saver,
config.mode == 'train')
utils.log()
if config.mode == 'train':
#summary_writer.add_graph(sess.graph)
utils.heading('START TRAINING ({:})'.format(config.model_name))
model_trainer.train(sess, progress, summary_writer)
elif config.mode == 'eval-train':
utils.heading('RUN EVALUATION ({:})'.format(config.model_name))
progress.best_model_saver.restore(sess, tf.train.latest_checkpoint(
config.checkpoints_dir))
model_trainer.evaluate_all_tasks(sess, summary_writer, None, train_set=True)
elif config.mode == 'eval-dev':
utils.heading('RUN EVALUATION ({:})'.format(config.model_name))
progress.best_model_saver.restore(sess, tf.train.latest_checkpoint(
config.checkpoints_dir))
model_trainer.evaluate_all_tasks(sess, summary_writer, None, train_set=False)
elif config.mode == 'infer':
utils.heading('START INFER ({:})'.format(config.model_name))
progress.best_model_saver.restore(sess, tf.train.latest_checkpoint(
config.checkpoints_dir))
model_trainer.infer(sess)
elif config.mode == 'translate':
utils.heading('START TRANSLATE ({:})'.format(config.model_name))
progress.best_model_saver.restore(sess, tf.train.latest_checkpoint(
config.checkpoints_dir))
model_trainer.translate(sess)
elif config.mode == 'eval-translate-train':
utils.heading('RUN EVALUATION ({:})'.format(config.model_name))
progress.best_model_saver.restore(sess, tf.train.latest_checkpoint(
config.checkpoints_dir))
model_trainer.evaluate_all_tasks(sess, summary_writer, None, train_set=True, is_translate=True)
elif config.mode == 'eval-translate-dev':
utils.heading('RUN EVALUATION ({:})'.format(config.model_name))
progress.best_model_saver.restore(sess, tf.train.latest_checkpoint(
config.checkpoints_dir))
model_trainer.evaluate_all_tasks(sess, summary_writer, None, train_set=False, is_translate=True)
else:
raise ValueError('Mode must be "train" or "eval"')