def report_training(self,
step,
num_steps,
learning_rate,
report_stats,
multigpu=False):
"""
This is the user-defined batch-level traing progress
report function.
Args:
step(int): current step count.
num_steps(int): total number of batches.
learning_rate(float): current learning rate.
report_stats(Statistics): old Statistics instance.
Returns:
report_stats(Statistics): updated Statistics instance.
"""
if self.start_time < 0:
raise ValueError("""ReportMgr needs to be started
(set 'start_time' or use 'start()'""")
if step % self.report_every == 0:
if multigpu:
report_stats = \
Statistics.all_gather_stats(report_stats)
self._report_training(step, num_steps, learning_rate, report_stats)
self.progress_step += 1
return Statistics()
else:
return report_stats
def _maybe_gather_stats(self, stat):
"""
Gather statistics in multi-processes cases
Args:
stat(:obj:onmt.utils.Statistics): a Statistics object to gather
or None (it returns None in this case)
Returns:
stat: the updated (or unchanged) stat object
"""
if stat is not None and self.n_gpu > 1:
return Statistics.all_gather_stats(stat)
return stat