def _end_epoch(self, epoch):
self._log_stats(epoch)
self.expl_data_collector.end_epoch(epoch)
ray.get([self.remote_eval_data_collector.end_epoch.remote(epoch)])
self.replay_buffer.end_epoch(epoch)
self.trainer.end_epoch(epoch)
# We can only save the state of the program
# after we call end epoch on all objects with internal state.
# This is so that restoring from the saved state will
# lead to identical result as if the program was left running.
snapshot = self._get_snapshot(epoch)
logger.save_itr_params(epoch + 1, snapshot)
gt.stamp('saving')
logger.record_dict(_get_epoch_timings())
logger.record_tabular('Epoch', epoch)
write_header = True if epoch == 0 else False
logger.dump_tabular(with_prefix=False,
with_timestamp=False,
write_header=write_header)
def _end_epoch(self, epoch):
self._log_stats(epoch)
if epoch > 0:
snapshot = self._get_snapshot(epoch)
logger.save_itr_params(epoch + 1, snapshot)
gt.stamp('saving', unique=False)
self.trainer.end_epoch(epoch)
logger.record_dict(_get_epoch_timings())
logger.record_tabular('Epoch', epoch)
write_header = True if epoch == 0 else False
logger.dump_tabular(with_prefix=False, with_timestamp=False,
write_header=write_header)
def new_add_scalar(key, value, step):
if isinstance(value, torch.Tensor):
value = ptu.get_numpy(value)
old_add_scalar(key, value, step)
logger.record_tabular(key, value)
