Unable to log_model="all" with wandb_logger

[robmarkcole]

Description

Configured to log all checkpoints, but only a single is logged

Steps to reproduce

from lightning.pytorch.loggers import WandbLogger

wandb_logger = WandbLogger(
    project="eurosat-test3", 
    name="imagenet", 
    log_model="all", 
    save_dir = "wandb_logs_tests"
)

trainer = Trainer(
    logger=wandb_logger,
    min_epochs=15,
    max_epochs=25,
)

trainer.fit(model=model, datamodule=datamodule)

glob("wandb_logs_tests/eurosat-test3/*/checkpoints/*.ckpt") # returns only a single checkpoint

Version

0.5.0

[calebrob6]

Is this a torchgeo bug or lightning bug?

[robmarkcole]

I commented out configure_callbacks(self) and now all logs are created by the callback:

However if I comment out the callback, I still only have 1 checkpoint (final) locally

However all checkpoints are uploaded

It is not clear but perhaps this is intended behaviour and not a bug. Either way I think this rules out this being a torchgeo bug so closing