def _build_cnn(self, units, n_hidden_list, cnn_filter_width,
use_batch_norm):
units = stacked_cnn(units,
n_hidden_list,
cnn_filter_width,
use_batch_norm,
training_ph=self.training_ph)
return units
def __init__(self,
word_vocab,
char_vocab,
tag_vocab,
n_filters=(128, 256),
filter_width=3,
token_embeddings_dim=128,
char_embeddings_dim=50,
use_char_embeddins=True,
embeddings_dropout=False,
dense_dropout=False,
use_batch_norm=False,
logging=False,
use_crf=False,
net_type='cnn',
char_filter_width=5,
verbose=False,
sess=None,
cell_type='lstm',
embedder=None,
use_capitalization=False,
two_dense_layers=False):
tf.set_random_seed(SEED)
n_tags = len(tag_vocab)
n_tokens = len(word_vocab)
n_chars = len(char_vocab)
# Create placeholders
if embedder is not None:
x_word = tf.placeholder(dtype=tf.float32,
shape=[None, None, token_embeddings_dim],
name='x_word')
else:
x_word = tf.placeholder(dtype=tf.int32,
shape=[None, None],
name='x_word')
x_char = tf.placeholder(dtype=tf.int32,
shape=[None, None, None],
name='x_char')
y_true = tf.placeholder(dtype=tf.int32,
shape=[None, None],
name='y_tag')
x_cap = tf.placeholder(dtype=tf.float32,
shape=[None, None],
name='y_cap')
# Auxiliary placeholders
learning_rate_ph = tf.placeholder(dtype=tf.float32,
shape=[],
name='learning_rate')
dropout_ph = tf.placeholder_with_default(1.0, shape=[])
training_ph = tf.placeholder_with_default(False, shape=[])
mask_ph = tf.placeholder(dtype=tf.float32, shape=[None, None])
# Embeddings
if embedder is None:
with tf.variable_scope('Embeddings'):
w_emb = embedding_layer(
x_word,
n_tokens=n_tokens,
token_embedding_dim=token_embeddings_dim)
if use_char_embeddins:
c_emb = character_embedding_network(
x_char,
n_characters=n_chars,
char_embedding_dim=char_embeddings_dim,
filter_width=char_filter_width)
emb = tf.concat([w_emb, c_emb], axis=-1)
else:
emb = w_emb
else:
emb = x_word
if use_capitalization:
emb = tf.concat([emb, tf.expand_dims(x_cap, 2)], 2)
# Dropout for embeddings
if embeddings_dropout:
emb = tf.layers.dropout(emb, dropout_ph, training=training_ph)
if 'cnn' in net_type.lower():
# Convolutional network
with tf.variable_scope('ConvNet'):
units = stacked_cnn(emb,
n_hidden_list=n_filters,
filter_width=filter_width,
use_batch_norm=use_batch_norm,
training_ph=training_ph)
elif 'rnn' in net_type.lower():
units, _ = stacked_bi_rnn(emb, n_filters, cell_type)
elif 'cnn_highway' in net_type.lower():
units = stacked_highway_cnn(emb,
n_hidden_list=n_filters,
filter_width=filter_width,
use_batch_norm=use_batch_norm,
training_ph=training_ph)
else:
raise KeyError(
'There is no such type of network: {}'.format(net_type))
# Classifier
with tf.variable_scope('Classifier'):
if two_dense_layers:
units = tf.layers.dense(
units,
n_filters[-1],
kernel_initializer=xavier_initializer())
logits = tf.layers.dense(units,
n_tags,
kernel_initializer=xavier_initializer())
# Loss with masking
if use_crf:
sequence_lengths = tf.reduce_sum(mask_ph, axis=1)
log_likelihood, trainsition_params = tf.contrib.crf.crf_log_likelihood(
logits, y_true, sequence_lengths)
loss_tensor = -log_likelihood
predictions = None
else:
ground_truth_labels = tf.one_hot(y_true, n_tags)
loss_tensor = tf.nn.softmax_cross_entropy_with_logits(
labels=ground_truth_labels, logits=logits)
loss_tensor = loss_tensor * mask_ph
predictions = tf.argmax(logits, axis=-1)
loss = tf.reduce_mean(loss_tensor)
# Initialize session
if sess is None:
sess = tf.Session()
if logging:
self.train_writer = tf.summary.FileWriter('summary', sess.graph)
self._token_embeddings_dim = token_embeddings_dim
self.token_vocab = word_vocab
self.tag_vocab = tag_vocab
self.char_vocab = char_vocab
self._use_crf = use_crf
self.summary = tf.summary.merge_all()
self._x_w = x_word
self._x_c = x_char
self._y_true = y_true
self._y_pred = predictions
self._x_cap = x_cap
if use_crf:
self._logits = logits
self._trainsition_params = trainsition_params
self._sequence_lengths = sequence_lengths
self._loss = loss
self._sess = sess
self._learning_rate_ph = learning_rate_ph
self._dropout = dropout_ph
self._loss_tensor = loss_tensor
self._use_dropout = True if embeddings_dropout or dense_dropout else None
self._training_ph = training_ph
self._logging = logging
self._train_op = self.get_train_op(loss, learning_rate_ph)
self._embedder = embedder
self.verbose = verbose
self._mask = mask_ph
self._use_capitalization = use_capitalization
sess.run(tf.global_variables_initializer())
def _build_cnn(self, units, n_hidden_list, cnn_filter_width, use_batch_norm):
units = stacked_cnn(units, n_hidden_list, cnn_filter_width, use_batch_norm, training_ph=self.training_ph)
return units