Change of behaviour from lightning logger about gluonts HOT 2 CLOSED

@cgoo4 estimators based on PyTorch use Lightning for training: the documentation for the latest version (https://lightning.ai/docs/pytorch/stable/extensions/logging.html) mentions the TensorboardLogger as being used by default. If I’m not wrong, older versions would use that if Tensorboard was installed, and the CVSLogger otherwise.

Things you could try:

• uninstalling Tensorboard from your environment
• explicitly specifying the CSVLogger as indicated in the docs above, via the trainer_kwargs to the gluonts estimator constructor.

(Removing the bug label)

from gluonts.

@lostella thank you for the detective work! The below code does bring back the metrics.csv. I must have installed Tensorboard as a dependency to something.

from lightning.pytorch.loggers import CSVLogger

logger = CSVLogger("logs")

trainer_kwargs = {"logger": logger}

from gluonts.