0
315
VistasHow do you prevent the tensorboard logger in pytorch lightning from logging the current epoch?
When creating a new tensorboard logger in pytorch lightning, the two things that are logged by default are the current epoch and the hp_metric. I was able to disable the hp_metric logging by setting default_hp_metric=False but I can't find anything to disable the logging of the epoch. I've searched in the lightning.py, trainer.py and tensorboard.py files which have the code for the module, the trainer and the tensorboard logger and couldn't find a logging call for epoch anywhere.
This behavior occurs even taking the barebones example from the pytorch lightning tutorial.
Is there a way to disable this logging of epoch to prevent clutter in the tensorboard interface?
1 Respuestas
Responde la pregunta0
In Short
You can disable automatically writing epoch variable by overwriting tensorboard logger.
from pytorch_lightning import loggers
from pytorch_lightning.utilities import rank_zero_only
class TBLogger(loggers.TensorBoardLogger):
@rank_zero_only
def log_metrics(self, metrics, step):
metrics.pop('epoch', None)
return super().log_metrics(metrics, step)
Full version
- Pytorch lightning automatically add
epochvsglobal_stepgraph to each logger. (you can see description in here) - There is no option to turn this behavior off. Because this is hard coded without any condition like below: (see full source code in here)
if step is None: # added metrics for convenience scalar_metrics.setdefault("epoch", self.trainer.current_epoch) step = self.trainer.global_step # log actual metrics self.trainer.logger.agg_and_log_metrics(scalar_metrics, step=step) - To disable this option, you should pop
epochvariable from metric dictionary inlog_metrics(metrics, step)that is called inadd_and_log_metrics(scalar_metrics, step=step). Code is shown in above. You can see full long version snippet in here.