def __init__(self, is_training=True, is_predict=False):
super(DiscriminantTrainer, self).__init__()
self.is_training = is_training
self.is_predict = is_predict
self.gen_text_feature = None
emb_dim = FLAGS.emb_dim
init_width = 0.5 / emb_dim
vocabulary.init()
vocab_size = vocabulary.get_vocab_size()
self.vocab_size = vocab_size
#if not cpu and on gpu run and using adagrad, will fail TODO check why
#also this will be more safer, since emb is large might exceed gpu mem
#with tf.device('/cpu:0'):
# #NOTICE if using bidirectional rnn then actually emb_dim is emb_dim / 2, because will at last step depth-concatate output fw and bw vectors
# self.emb = melt.variable.get_weights_uniform('emb', [vocab_size, emb_dim], -init_width, init_width)
self.emb = embedding.get_embedding_cpu('emb')
melt.visualize_embedding(self.emb, FLAGS.vocab)
if is_training and FLAGS.monitor_level > 0:
melt.monitor_embedding(self.emb, vocabulary.vocab, vocab_size)
self.activation = melt.activations[FLAGS.activation]
self.weights_initializer = tf.random_uniform_initializer(
-FLAGS.initializer_scale, FLAGS.initializer_scale)
self.biases_initialzier = melt.slim.init_ops.zeros_initializer if FLAGS.bias else None
if not FLAGS.pre_calc_image_feature:
assert melt.apps.image_processing.image_processing_fn is not None, 'forget melt.apps.image_processing.init()'
self.image_process_fn = functools.partial(
melt.apps.image_processing.image_processing_fn,
height=FLAGS.image_height,
width=FLAGS.image_width)
def __init__(self, encoder_type='bow', is_training=True, is_predict=False):
super(DualTextsim, self).__init__()
self.is_training = is_training
self.is_predict = is_predict
self.encoder = encoder_factory.get_encoder(encoder_type, is_training,
is_predict)
self.encoder_type = encoder_type
emb_dim = FLAGS.emb_dim
init_width = 0.5 / emb_dim
vocabulary.init()
vocab_size = vocabulary.get_vocab_size()
self.vocab_size = vocab_size
# 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 finetune=:{}'
.format(FLAGS.model_dir, FLAGS.finetune_word_embedding))
self.emb = embedding.get_embedding_cpu(
name='emb', trainable=FLAGS.finetune_word_embedding)
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 finetune=:{}'.format(
FLAGS.word_embedding_file, FLAGS.finetune_word_embedding))
self.emb = melt.load_constant_cpu(
FLAGS.word_embedding_file,
name='emb',
trainable=FLAGS.finetune_word_embedding)
if FLAGS.position_embedding:
logging.info('Using position embedding')
self.pos_emb = embedding.get_embedding_cpu(name='pos_emb',
height=TEXT_MAX_WORDS)
else:
self.pos_emb = None
melt.visualize_embedding(self.emb, FLAGS.vocab)
if is_training and FLAGS.monitor_level > 0:
melt.monitor_embedding(self.emb, vocabulary.vocab, vocab_size)
self.activation = melt.activations[FLAGS.activation]
# TODO can consider global initiallizer like
# with tf.variable_scope("Model", reuse=None, initializer=initializer)
# https://github.com/tensorflow/models/blob/master/tutorials/rnn/ptb/ptb_word_lm.py
self.weights_initializer = tf.random_uniform_initializer(
-FLAGS.initializer_scale, FLAGS.initializer_scale)
self.biases_initialzier = melt.slim.init_ops.zeros_initializer if FLAGS.bias else None
self.mlp_dims = [int(x) for x in FLAGS.mlp_dims.split(',')
] if FLAGS.mlp_dims is not '0' else None
self.scope = 'dual_textsim'
self.build_train_graph = self.build_graph
def __init__(self, is_training=False, is_predict=False,
emb_dim=None, initializer=None):
ShowAndTellEncoder.__init__(self, is_training, is_predict, emb_dim, initializer)
self.pos_emb = embedding.get_embedding_cpu(name='pos_emb', height=FLAGS.image_attention_size)
def __init__(self, encoder_type='bow', is_training=True, is_predict=False):
super(DiscriminantTrainer, self).__init__()
self.is_training = is_training
self.is_predict = is_predict
logging.info('emb_dim:{}'.format(FLAGS.emb_dim))
logging.info('margin:{}'.format(FLAGS.margin))
self.encoder = encoder_factory.get_encoder(encoder_type, is_training,
is_predict)
self.encoder_type = encoder_type
emb_dim = FLAGS.emb_dim
init_width = 0.5 / emb_dim
vocabulary.init()
vocab_size = vocabulary.get_vocab_size()
self.vocab_size = vocab_size
#if not cpu and on gpu run and using adagrad, will fail TODO check why
#also this will be more safer, since emb is large might exceed gpu mem
#with tf.device('/cpu:0'):
# self.emb = melt.variable.get_weights_uniform('emb', [vocab_size, emb_dim], -init_width, init_width)
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 finetune=:{}'
.format(FLAGS.model_dir, FLAGS.finetune_word_embedding))
self.emb = embedding.get_embedding_cpu(
name='emb', trainable=FLAGS.finetune_word_embedding)
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 finetune=:{}'.format(
FLAGS.word_embedding_file, FLAGS.finetune_word_embedding))
self.emb = melt.load_constant_cpu(
FLAGS.word_embedding_file,
name='emb',
trainable=FLAGS.finetune_word_embedding)
melt.visualize_embedding(self.emb, FLAGS.vocab)
if is_training and FLAGS.monitor_level > 0:
melt.monitor_embedding(self.emb, vocabulary.vocab, vocab_size)
self.activation = melt.activations[FLAGS.activation]
#TODO can consider global initiallizer like
# with tf.variable_scope("Model", reuse=None, initializer=initializer)
#https://github.com/tensorflow/models/blob/master/tutorials/rnn/ptb/ptb_word_lm.py
self.weights_initializer = tf.random_uniform_initializer(
-FLAGS.initializer_scale, FLAGS.initializer_scale)
self.biases_initialzier = melt.slim.init_ops.zeros_initializer if FLAGS.bias else None
if not FLAGS.pre_calc_image_feature:
assert melt.apps.image_processing.image_processing_fn is not None, 'forget melt.apps.image_processing.init()'
self.image_process_fn = functools.partial(
melt.apps.image_processing.image_processing_fn,
height=FLAGS.image_height,
width=FLAGS.image_width)
self.image_mlp_dims = [
int(x) for x in FLAGS.image_mlp_dims.split(',')
] if FLAGS.image_mlp_dims is not '0' else None
self.text_mlp_dims = [int(x) for x in FLAGS.text_mlp_dims.split(',')
] if FLAGS.text_mlp_dims is not '0' else None
self.scope = 'image_text_sim'