def forward_fixed_text(self, text_npy):
#text = tf.constant(text_npy) #smaller than 2G, then ok...
#but for safe in more application
text = melt.load_constant(text_npy, self.sess)
with tf.variable_scope(self.scope):
text_feature = self.forward_text(text)
return text_feature
def __init__(self,
emb=None,
fixed_emb=None,
name=None,
fixed_name=None,
model_dir=None,
model_name=None,
sess=None):
self._sess = sess or tf.InteractiveSession()
if os.path.isdir(emb):
model_dir = emb
if model_dir is None:
if isinstance(emb, str):
emb = np.load(emb)
emb = melt.load_constant(emb, name=name)
if isinstance(fixed_emb, str):
#fixed_emb is corpus embeddings, all sumed and normed already
fixed_emb = melt.load(fixed_emb, name=fixed_emb)
else:
model_path = melt.get_model_path(model_dir, model_name)
emb = tf.Variable(0., name=name, validate_shape=False)
#emb = tf.Variable(0., name=name)
#like word2vec the name is 'w_in'
embedding_saver = tf.train.Saver({name: emb})
embedding_saver.restore(self._sess, model_path)
#assume 0 index not used, 0 for PAD
mask = zero_first_row(emb)
emb = tf.multiply(emb, mask)
self._emb = emb
self._fixed_emb = fixed_emb
self._normed_emb = None
def init_evaluate_constant_text(self, text_npy):
#self.text = tf.constant(text_npy)
if self.text is None:
#if self.encoder_type == 'cnn':
# import numpy as np
# #text_npy = np.load(text_npy)
# text_npy = text_npy[:5000]
# #self.text = melt.load_constant_cpu(text_npy, self.sess)
##else:
self.text = melt.load_constant(text_npy, self.sess)
def get_or_restore_embedding(name='emb',
embedding_file=None,
trainable=None,
height=None,
emb_dim=None,
type='word'):
# cpu for adgrad optimizer
#if (not FLAGS.word_embedding_file) or glob.glob(FLAGS.model_dir + '/model*ckpt*'):
# logging.info('Word embedding random init or from model_dir:{} and trainable=:{}'.format(
# FLAGS.model_dir, FLAGS.finetune_word_embedding))
#TODO verify below is ok , above is ok but a bit complex. I assume if var in check point will later restore and cover initital const value
#if not FLAGS.word_embedding_file:
embedding_file_ = None
train_able_ = None
if type == 'word':
embedding_file_ = FLAGS.word_embedding_file
train_able_ = FLAGS.finetune_word_embedding
elif type == 'char':
embedding_file_ = FLAGS.char_embedding_file
train_able_ = FLAGS.finetune_char_embedding
elif type == 'ngram':
embedding_file_ = FLAGS.ngram_embedding_file
train_able_ = FLAGS.finetune_ngram_embedding
elif type == 'pinyin':
embedding_file_ = FLAGS.pinyin_embedding_file
train_able_ = FLAGS.finetune_pinyin_embedding
else:
raise ValueError(type)
embedding_file = embedding_file if embedding_file is not None else embedding_file_
trainable = trainable if trainable is not None else train_able_
#logging.info('----------------------', type, embedding_file, height)
if (not embedding_file) or melt.exists_model(FLAGS.model_dir):
logging.info(
'{} random init or from model_dir and trainable=:{}'.format(
name, trainable))
emb = get_embedding(name=name,
trainable=trainable,
height=height,
emb_dim=emb_dim)
#melt.try_add_to_collection('word_embedding', emb)
else:
# https://github.com/tensorflow/tensorflow/issues/1570
# still adgrad must cpu..
# if not fintue emb this will be ok if fintune restart will ok ? must not use word embedding file? os.path.exists(FLAGS.model_dir) ? judge?
# or will still try to load from check point ? TODO for safe you could re run by setting word_embedding_file as None or ''
logging.info('Loading {} from:{} and trainable=:{}'.format(
name, embedding_file, trainable))
timer = gezi.Timer('load constat')
emb = melt.load_constant(embedding_file,
name=name,
trainable=trainable)
timer.print_elapsed()
return emb
def build_fixed_text_feature_graph(self, text_feature_npy):
"""
text features directly load to graph, @NOTICE text_feature_npy all vector must of same length
used in evaluate.py for both fixed text and fixed words
@FIXME dump text feature should change api
"""
with tf.variable_scope("image_text_sim"):
image_feature = self.forward_image_feature(self.image_feature_feed)
text_feature = melt.load_constant(self.sess, text_feature_npy)
score = melt.cosine(image_feature, text_feature, nonorm=True)
return score
def get_or_restore_embedding(name='emb'):
# cpu for adgrad optimizer
if (not FLAGS.word_embedding_file) or glob.glob(FLAGS.model_dir +
'/model.ckpt*'):
logging.info(
'Word embedding random init or from model_dir:{} and trainable=:{}'
.format(FLAGS.model_dir, FLAGS.finetune_word_embedding))
emb = get_embedding(name=name, trainable=FLAGS.finetune_word_embedding)
melt.try_add_to_collection('word_embedding', emb)
else:
# https://github.com/tensorflow/tensorflow/issues/1570
# still adgrad must cpu..
# if not fintue emb this will be ok if fintune restart will ok ? must not use word embedding file? os.path.exists(FLAGS.model_dir) ? judge?
# or will still try to load from check point ? TODO for safe you could re run by setting word_embedding_file as None or ''
logging.info('Loading word embedding from:{} and trainable=:{}'.format(
FLAGS.word_embedding_file, FLAGS.finetune_word_embedding))
emb = melt.load_constant(FLAGS.word_embedding_file,
name=name,
trainable=FLAGS.finetune_word_embedding)
return emb
def init_evaluate_constant_text(self, text_npy):
#self.text = tf.constant(text_npy)
if self.text is None:
self.text = melt.load_constant(self.sess, text_npy)