def _do_evaluation(self, run_context, global_step):
""" Evaluates the model by loss function.
Furthermore, if the decay_type of optimizer is "loss_decay", anneal
the learning rate at the right time.
Args:
run_context: A `SessionRunContext` object.
global_step: A python integer, the current training step.
"""
loss = evaluate(sess=run_context.session,
eval_op=self._loss_op,
feeding_data=self._eval_feeding_data)
tf.logging.info("Evaluating DEVSET: DevLoss=%f GlobalStep=%d" %
(loss, global_step))
if self._summary_writer is not None:
self._summary_writer.add_summary("Metrics/DevLoss", loss,
global_step)
if self._half_lr:
if loss <= self._min_loss:
self._min_loss = loss
self._patience = 0
else:
self._patience += 1
if self._patience >= self._max_patience:
self._patience = 0
run_context.session.run(self._half_lr_op)
now_lr = run_context.session.run(self._learning_rate)
tf.logging.info(
"Hit maximum patience=%d. HALF THE LEARNING RATE TO %f at %d"
% (self._max_patience, now_lr, global_step))
def _do_evaluation(self, run_context, global_step):
""" Evaluates the model by loss function.
Furthermore, if the decay_type of optimizer is "loss_decay", anneal
the learning rate at the right time.
Args:
run_context: A `SessionRunContext` object.
global_step: A python integer, the current training step.
"""
loss = evaluate(sess=run_context.session,
loss_op=self._loss_op,
eval_data=self._eval_feeding_data)
tf.logging.info("Evaluating DEVSET: DevLoss=%f GlobalStep=%d" % (loss, global_step))
if self._summary_writer is not None:
self._summary_writer.add_summary("Metrics/DevLoss", loss, global_step)
if self._half_lr and global_step >= self._start_decay_at:
if loss <= self._min_loss:
self._min_loss = loss
self._bad_count = 0
else:
self._bad_count += 1
if self._bad_count >= self._max_patience:
self._bad_count = 0
run_context.session.run(self._half_lr_op)
now_lr = run_context.session.run(self._learning_rate)
tf.logging.info("Hit maximum patience=%d. HALF THE LEARNING RATE TO %f at %d"
% (self._max_patience, now_lr, global_step))