Loggers in PyTorch Lightning

PyTorch Lightning has a builtin logger that can be used to log metrics and hyperparameters. Internally, it uses TensorBoard. However, it has a few caveats

• Logging is buggy; I couldn’t use it to log accuracy in the _epoch_end method
• It doesn’t support advanced functionality such as computational graphs
• There is no way to control the x-axis parameter

There are two workarounds

(1) Use TensorBoardLogger

This explicitly uses TensorBoard.

First you need to import it

from pytorch_lightning.loggers import TensorBoardLogger

It needs to be passed to the trainer. Then, inside the training/validation steps, you can log metrics by calling the method self.logger.experiment.add_scalar. Note that the parameter on the x-axis can also be specified.

(2) Use CSVLogger

This logs to CSV files, which can be used to plot data using matplotlib (for old-fashioned people like me!).

First you need to import it

from pytorch_lightning.loggers import CSVLogger

It needs to be passed to the trainer. Then, inside the training/validation steps, you can log metrics by calling the method self.logger.log_metrics

For complete working example on the MNIST dataset, please look at the notebooks here.