def __init__(self, opts, vocab,
classes): # opts = tf.app.flags.FLAGS object
self.opt = opts
self.vocab = vocab
self.classes = classes
batch_size = self.opt.batch_size
dropout_rate = self.opt.dropout_rate
glove_embeddings = tf.constant(self.vocab.embeddings(),
dtype=tf.float32)
self.emb = tf.get_variable("embeddings", initializer=glove_embeddings)
# create placeholders
self.encoder_inputs = tf.placeholder(shape=[batch_size, None],
dtype=tf.int32,
name='encoder_inputs')
self.encoder_input_lengths = tf.placeholder(
shape=[batch_size], dtype=tf.int32, name='encoder_input_lengths')
self.is_training = tf.placeholder(shape=[],
dtype=tf.bool,
name='is_training')
self.target_labels = tf.placeholder(shape=[batch_size],
dtype=tf.int32,
name='target_labels')
self.keep_prob = get_keep_prob(dropout_rate, self.is_training)
self.graph_built = False
self.train_op_added = False
def __init__(self, opts, vocab): # opts = tf.app.flags.FLAGS object
self.opt = opts
self.vocab = vocab
batch_size = self.opt.batch_size
dropout_rate = self.opt.dropout_rate
self.emb = tf.get_variable(
"embeddings",
shape=[vocab.size(), vocab.emb_size()],
initializer=tf.contrib.layers.xavier_initializer())
# create placeholders
self.encoder_inputs = tf.placeholder(shape=[batch_size, None],
dtype=tf.int32,
name='encoder_inputs')
self.encoder_input_lengths = tf.placeholder(
shape=[batch_size], dtype=tf.int32, name='encoder_input_lengths')
self.decoder_targets = tf.placeholder(shape=[batch_size, None],
dtype=tf.int32,
name='decoder_targets')
self.decoder_target_lengths = tf.placeholder(
shape=[batch_size], dtype=tf.int32, name='decoder_target_lengths')
self.decoder_inputs = tf.placeholder(shape=[batch_size, None],
dtype=tf.int32,
name='decoder_inputs')
self.decoder_input_lengths = tf.placeholder(
shape=[batch_size], dtype=tf.int32, name='decoder_input_lengths')
self.max_decoder_length = tf.placeholder(shape=[],
dtype=tf.int32,
name='max_decoder_length')
self.max_encoder_length = tf.placeholder(shape=[],
dtype=tf.int32,
name='max_encoder_length')
self.is_training = tf.placeholder(shape=[],
dtype=tf.bool,
name='is_training')
self.memory_vectors = tf.placeholder(shape=[batch_size, None],
dtype=tf.int32,
name='memory_vectors')
self.memory_vector_lengths = tf.placeholder(
shape=[batch_size], dtype=tf.int32, name='memory_vector_lengths')
self.oov_size = tf.placeholder(shape=[],
dtype=tf.int32,
name='oov_size')
self.memory_vectors_full_vocab = tf.placeholder(
shape=[batch_size, None],
dtype=tf.int32,
name='memory_vectors_full_vocab')
self.memory_elmo_token_ids = tf.placeholder(shape=[batch_size, None],
dtype=tf.int32)
self.query_elmo_token_ids = tf.placeholder(shape=[batch_size, None],
dtype=tf.int32)
self.elmo_batcher = TokenBatcher(self.opt.elmo_vocab_file)
self.keep_prob = get_keep_prob(dropout_rate, self.is_training)
self.oov_dicts = Queue()
self.graph_built = False
self.train_op_added = False
self.elmo_bilm = BidirectionalLanguageModel(
self.opt.elmo_options_file,
self.opt.elmo_weight_file,
use_character_inputs=False,
embedding_weight_file=self.opt.elmo_token_embedding_file)