def update_metainfo_op_with_vars(metainfo_ph: tf.Tensor, nz_ph: tf.Tensor,
metainfo_var: tf.Variable,
nz_var: tf.Variable) -> tf.Operation:
assign_nz = nz_var.assign(nz_ph)
assign_meta = metainfo_var.assign(metainfo_ph)
with tf.control_dependencies([assign_nz, assign_meta]):
update_op = tf.no_op()
return update_op
def _train_eval(self):
self.mask = tf.sequence_mask(self.features[constants.LENGTH_KEY],
name="padding_mask")
num_labels = self.extractor.vocab_size()
_logits = select_logits(self.logits, self.predicate_indices,
self.n_steps)
seq_mask = None if constants.BERT_LENGTH_KEY in self.features else self.features.get(
constants.SEQUENCE_MASK)
rel_loss = sequence_loss(
logits=_logits,
targets=self.targets,
sequence_lengths=self._sequence_lengths,
num_labels=num_labels,
crf=self.config.crf,
tag_transitions=self._tag_transitions,
label_smoothing=self.config.label_smoothing,
confidence_penalty=self.config.confidence_penalty,
name="bilinear_loss",
mask=seq_mask)
self.loss = rel_loss
self.metric = Variable(0,
name=append_label(constants.OVERALL_KEY,
self.name),
dtype=tf.float32,
trainable=False)
def _get_update(self, variable: tf.Variable, gradient: tf.Tensor,
step_size: tf.Tensor) -> List[tf.Operation]:
with tf.variable_scope(variable.op.name):
gradient = tf.cast(gradient, tf.float32)
state_m = tf.get_variable(
"adam_m",
shape=variable.shape,
dtype=variable.dtype,
initializer=tf.zeros_initializer(),
trainable=False,
)
updated_m = (self.beta1 * tf.cast(state_m, tf.float32) +
(1 - self.beta1) * gradient)
state_v = tf.get_variable(
"adam_v",
shape=variable.shape,
dtype=tf.float32,
initializer=tf.zeros_initializer(),
trainable=False,
)
updated_v = self.beta2 * state_v + (1 - self.beta2) * (gradient**2)
delta = step_size * updated_m / (tf.sqrt(updated_v) + self.epsilon)
updated_variable = tf.cast(variable, tf.float32) - delta
return [
variable.assign(tf.cast(updated_variable, variable.dtype)),
state_m.assign(tf.cast(updated_m, state_m.dtype)),
state_v.assign(updated_v),
]
def update_metainfo_op_with_vars(metainfo_ph: tf.Tensor, nz_ph: tf.Tensor,
metainfo_var: tf.Variable,
nz_var: tf.Variable) -> tf.Operation:
"""Returns an op that can be used to update the metainfo on device
:param metainfo_ph: Metainfo placeholder
:param nz_ph: Nonzero-values placeholder
:param metainfo_var: Metainfo variable
:param nz_var: Nonzero-values variable
"""
assign_nz = nz_var.assign(nz_ph)
assign_meta = metainfo_var.assign(metainfo_ph)
with tf.control_dependencies([assign_nz, assign_meta]):
update_op = tf.no_op()
return update_op
def _train_eval(self):
if self.config.label_smoothing > 0:
targets = tf.one_hot(self.targets,
depth=self.extractor.vocab_size())
self.loss = tf.losses.softmax_cross_entropy(
onehot_labels=targets,
logits=self.logits,
label_smoothing=self.config.label_smoothing)
else:
self.loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=self.logits, labels=self.targets))
self.metric = Variable(0,
name=append_label(constants.OVERALL_KEY,
self.name),
dtype=tf.float32,
trainable=False)
def _train_eval(self):
num_labels = self.extractor.vocab_size()
seq_mask = None if constants.BERT_LENGTH_KEY in self.features else self.features.get(
constants.SEQUENCE_MASK)
self.loss = sequence_loss(
logits=self.logits,
targets=self.targets,
sequence_lengths=self._sequence_lengths,
num_labels=num_labels,
crf=self.config.crf,
tag_transitions=self._tag_transitions,
label_smoothing=self.config.label_smoothing,
confidence_penalty=self.config.confidence_penalty,
mask=seq_mask)
self.metric = Variable(0,
name=append_label(constants.OVERALL_KEY,
self.name),
dtype=tf.float32,
trainable=False)
def _train_eval(self):
self.mask = tf.sequence_mask(self.lens, name="padding_mask")
# compute combined arc and rel losses (both via softmax cross entropy)
def compute_loss(logits, targets, name):
with tf.variable_scope(name):
losses = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=targets)
losses = tf.boolean_mask(losses, self.mask)
return tf.reduce_mean(losses)
self.arc_targets = tf.identity(self.features[constants.HEAD_KEY], name=constants.HEAD_KEY)
arc_loss = compute_loss(self.arc_logits, self.arc_targets, "arc_bilinear_loss")
_rel_logits = select_logits(self.rel_logits, self.arc_targets, self.n_steps)
rel_loss = compute_loss(_rel_logits, self.targets, "rel_bilinear_loss")
arc_loss = self.config.get('arc_loss_weight', 1) * arc_loss
rel_loss = self.config.get('rel_loss_weight', 1) * rel_loss
self.loss = arc_loss + rel_loss
self.metric = Variable(0, name=append_label(constants.OVERALL_KEY, self.name), dtype=tf.float32, trainable=False)