def _get_default_feed_dict(self, batch, is_training):
"""Reset status if necessary"""
feed_dict = super()._get_default_feed_dict(batch, is_training)
assert isinstance(batch, DataSet)
# (1) If a new sequence begin during training or validation, reset state
reset_flag = batch.should_reset_state
if hub.supreme_reset_flag is not None:
reset_flag = hub.supreme_reset_flag
if reset_flag:
self.reset_buffers(batch.size, is_training)
if is_training and hub.notify_when_reset: console.write_line('- ' * 40)
# (2) Decrease batch size if necessary
if batch.active_indices is not None:
# This signal is set in DataSet.gen_rnn_batches invoked by SequenceSet
self.decrease_buffer_size(batch.active_indices, is_training)
# BETA: for parallel engine logic
if batch.should_partially_reset_state:
self.reset_part_buffer(batch.reset_batch_indices, batch.reset_values)
# TODO: to be deprecated
if hub.notify_when_reset and False:
if batch.reset_values is not None:
info = [(i, v) for i, v in zip(
batch.reset_batch_indices, batch.reset_values)]
else: info = batch.reset_batch_indices
console.write_line('{}'.format(info))
# (3) Set status buffer to status placeholder
feed_dict.update(self._get_rnn_dict(is_training, batch.size))
return feed_dict
def end_round(self, rnd):
for i, metric in enumerate(self._metrics):
assert isinstance(metric, Metric) and metric.activated
if metric.sleep: continue
console.write_line('Branch {} {}'.format(i + 1, '- ' * 35))
metric.end_round(rnd)
console.write_line('- ' * 40)
def _get_default_feed_dict(self, batch, is_training):
feed_dict = Feedforward._get_default_feed_dict(self, batch,
is_training)
if self.master is Recurrent:
assert isinstance(batch, DataSet)
# If a new sequence begin while training, reset state
if is_training:
if batch.should_reset_state:
if hub.notify_when_reset: console.write_line('- ' * 40)
self.reset_state(batch.size)
if batch.should_partially_reset_state:
if hub.notify_when_reset:
if batch.reset_values is not None:
info = [(i, v) for i, v in zip(
batch.reset_batch_indices, batch.reset_values)]
else:
info = batch.reset_batch_indices
console.write_line('{}'.format(info))
self.reset_part_state(batch.reset_batch_indices,
batch.reset_values)
batch_size = None if is_training else batch.size
# If is not training, always set a zero state to model
feed_dict.update(self._get_state_dict(batch_size=batch_size))
return feed_dict
def _snapshot(self, progress):
if self._snapshot_function is None:
return
filename = 'train_{}_episode'.format(self.counter)
fullname = "{}/{}".format(self.snapshot_dir, filename)
self._snapshot_function(fullname)
console.clear_line()
console.write_line("[Snapshot] snapshot saved to {}".format(filename))
console.print_progress(progress=progress)
def show_notes(self):
console.section('Notes')
console.write_line(self._note.content)