Open
Description
Description & Motivation
Periodic flushing of Comet experiment data by CometLogger
is sometimes necessary to ensure all metrics and data remain up to date in the Comet UI and aligned with saved model checkpoints. Some users accomplish this by creating custom Callbacks, though this approach can be rather cumbersome.
Pitch
It is proposed to enhance the CometLogger
implementation to support flushing collected data to Comet after each checkpoint is saved. This would be configurable via the flush_every
parameter of the CometLogger
.
This improvement allows users to manage Comet experiment flushing using the standard PyTorch Lightning approach by registering appropriate ModelCheckpoint
callbacks with the Trainer
.
Alternatives
No response
Additional context
No response