def bag_cross_max(x,
scope,
rel_tot,
is_training,
var_scope=None,
dropout_before=False,
keep_prob=1.0):
"""
Cross-sentence Max-pooling proposed by (Jiang et al. 2016.)
"Relation Extraction with Multi-instance Multi-label Convolutional Neural Networks"
https://pdfs.semanticscholar.org/8731/369a707046f3f8dd463d1fd107de31d40a24.pdf
"""
with tf.variable_scope(var_scope or "cross_max", reuse=tf.AUTO_REUSE):
if dropout_before:
x = dropout(x, keep_prob)
bag_repre = []
for i in range(scope.shape[0] - 1):
bag_hidden_mat = x[scope[i]:scope[i + 1]]
bag_repre.append(tf.reduce_max(
bag_hidden_mat, 0)) # (n', hidden_size) -> (hidden_size)
bag_repre = tf.stack(bag_repre)
if not dropout_before:
bag_repre = dropout(bag_repre, keep_prob)
bag_logit = _logit(bag_repre, rel_tot)
if not is_training:
bag_logit = tf.nn.softmax(bag_logit)
return bag_logit, bag_repre
def bag_attention(x,
scope,
instance_label,
rel_tot,
is_training,
var_scope=None,
dropout_before=False,
keep_prob=1.0):
with tf.variable_scope(var_scope or "bag_attention", reuse=tf.AUTO_REUSE):
if is_training: # training
if dropout_before:
x = dropout(x, keep_prob)
bag_repre = []
attention_logit = _attention_train_logit(x, instance_label,
rel_tot)
for i in range(scope.shape[0] - 1):
bag_hidden_mat = x[scope[i]:scope[i + 1]]
attention_score = tf.nn.softmax(
attention_logit[scope[i]:scope[i + 1]], -1)
# (1, n') x (n', hidden_size) = (1, hidden_size) -> (hidden_size)
bag_repre.append(
tf.squeeze(
tf.matmul(tf.expand_dims(attention_score, 0),
bag_hidden_mat)))
bag_repre = tf.stack(bag_repre)
if not dropout_before:
bag_repre = dropout(bag_repre, keep_prob)
return _logit(bag_repre, rel_tot), bag_repre
else: # testing
attention_logit = _attention_test_logit(x, rel_tot) # (n, rel_tot)
bag_repre = []
bag_logit = []
for i in range(scope.shape[0] - 1):
bag_hidden_mat = x[scope[i]:scope[i + 1]]
attention_score = tf.nn.softmax(
tf.transpose(attention_logit[scope[i]:scope[i + 1], :]),
-1) # softmax of (rel_tot, n')
bag_repre_for_each_rel = tf.matmul(
attention_score, bag_hidden_mat
) # (rel_tot, n') \dot (n', hidden_size) = (rel_tot, hidden_size)
bag_logit_for_each_rel = _logit(
bag_repre_for_each_rel, rel_tot) # -> (rel_tot, rel_tot)
bag_repre.append(bag_repre_for_each_rel)
bag_logit.append(
tf.diag_part(
tf.nn.softmax(bag_logit_for_each_rel,
-1))) # could be improved by sigmoid?
bag_repre = tf.stack(bag_repre)
bag_logit = tf.stack(bag_logit)
return bag_logit, bag_repre
def cnn(x, mask=None, hidden_size=230, kernel_size=3, stride_size=1, activation=tf.nn.relu,
var_scope=None, keep_prob=1.0):
with tf.variable_scope(var_scope or ('cnn' if mask is None else 'pcnn'), reuse=tf.AUTO_REUSE):
cnn_cell = _cnn_cell(x, hidden_size, kernel_size, stride_size)
pool = _pooling(cnn_cell) if mask is None else _piecewise_pooling(cnn_cell, mask)
return dropout(activation(pool), keep_prob)
def bag_one(x,
scope,
label,
rel_tot,
is_training,
var_scope=None,
dropout_before=False,
keep_prob=1.0):
with tf.variable_scope(var_scope or "maximum", reuse=tf.AUTO_REUSE):
if is_training: # training
if dropout_before:
x = dropout(x, keep_prob)
bag_repre = []
for i in range(scope.shape[0] - 1):
bag_hidden_mat = x[scope[i]:scope[i + 1]]
instance_logit = tf.nn.softmax(
_logit(bag_hidden_mat,
rel_tot), -1) # (n', hidden_size)->(n', rel_tot)
j = tf.argmax(instance_logit[:, label[i]],
output_type=tf.int32)
bag_repre.append(bag_hidden_mat[j])
bag_repre = tf.stack(bag_repre)
if not dropout_before:
bag_repre = dropout(bag_repre, keep_prob)
return _logit(bag_repre, rel_tot), bag_repre
else: # testing
if dropout_before:
x = dropout(x, keep_prob)
bag_repre = []
bag_logit = []
for i in range(scope.shape[0] - 1):
bag_hidden_mat = x[scope[i]:scope[i + 1]]
instance_logit = tf.nn.softmax(
_logit(bag_hidden_mat,
rel_tot), -1) # (n', hidden_size)->(n', rel_tot)
bag_logit.append(tf.reduce_max(instance_logit, 0))
bag_repre.append(bag_hidden_mat[0]) # fake max repre
bag_logit = tf.stack(bag_logit)
bag_repre = tf.stack(bag_repre)
return tf.nn.softmax(bag_logit), bag_repre
def rnn(x, length, hidden_size=230, cell_name='', bidirectional=False, var_scope=None, keep_prob=1.0):
with tf.variable_scope(var_scope or ('birnn' if bidirectional else 'rnn'), reuse=tf.AUTO_REUSE):
x = dropout(x, keep_prob)
if bidirectional:
bw_states, fw_states = birnn_states(x, length, hidden_size, cell_name)
return tf.concat([fw_states, bw_states], axis=1)
else:
cell = _rnn_cell(hidden_size, cell_name)
_, states = tf.nn.dynamic_rnn(cell, x, sequence_length=length, dtype=tf.float32, scope='dynamic_rnn')
if isinstance(states, tuple):
states = states[0]
return states
def bag_average(x,
scope,
rel_tot,
is_training,
var_scope=None,
dropout_before=False,
keep_prob=1.0):
with tf.variable_scope(var_scope or "average", reuse=tf.AUTO_REUSE):
if dropout_before:
x = dropout(x, keep_prob)
bag_repre = []
for i in range(scope.shape[0] - 1):
bag_hidden_mat = x[scope[i]:scope[i + 1]]
bag_repre.append(tf.reduce_mean(
bag_hidden_mat, 0)) # (n', hidden_size) -> (hidden_size)
bag_repre = tf.stack(bag_repre)
if not dropout_before:
bag_repre = dropout(bag_repre, keep_prob)
bag_logit = _logit(bag_repre, rel_tot)
if not is_training:
bag_logit = tf.nn.softmax(bag_logit)
return bag_logit, bag_repre
def resnet(x, filters, length=None, cell_name='lstm', bidirectional=False, mask=None, ib_num=4,
kernel_size=3, stride_size=1, activation=tf.nn.relu, var_scope=None, keep_prob=1.0):
with tf.variable_scope(var_scope or ('resnet' if mask is None else 'resnet_pcnn'), reuse=tf.AUTO_REUSE):
seq = None if length is None else rnn(x, length, filters[1], cell_name, bidirectional, keep_prob=keep_prob)
x = _cnn_cell(x, filters[1], kernel_size, stride_size, activation=activation)
# x = tf.expand_dims(_pooling(x), axis=1) if mask is None else _piecewise_pooling(x, mask, True)
for i in range(ib_num):
h1 = _cnn_cell(x, filters[0], kernel_size, stride_size, activation=activation,
var_scope='conv_' + str(i) + 'a')
h2 = _cnn_cell(h1, filters[1], kernel_size, stride_size, activation=activation,
var_scope='conv_' + str(i) + 'b')
x = h2 + x
x = _pooling(x) if mask is None else _piecewise_pooling(x, mask)
# x = tf.squeeze(x) if mask is None else tf.reshape(x, [-1, x.shape[-1] * x.shape[-2]])
# x = conv_block(x, kernel_size, [hidden_size, hidden_size, 256], stage=2, block='a',
# strides=(stride_size, stride_size))
# x = identity_block(x, kernel_size, [hidden_size, hidden_size, 256], stage=2, block='b')
# x = identity_block(x, kernel_size, [hidden_size, hidden_size, 256], stage=2, block='c')
x = dropout(activation(x), keep_prob)
return x if seq is None else tf.concat([seq, x], axis=1)
def instance(x, rel_tot, var_scope=None, keep_prob=1.0):
with tf.variable_scope(var_scope or "instance", reuse=tf.AUTO_REUSE):
x = dropout(x, keep_prob)
return _logit(x, rel_tot), x