mlflow checkpoints in the wrong location

[david-waterworth]

I'm not sure if I'm doing something wrong, I'm using mlflow instead of tensorboard as a logger. I've used the defaults i.e.

mlflow = loggers.MLFlowLogger()
trainer = pl.Trainer.from_argparse_args(args, logger=mlflow)

I'm ending up with the following folder structure

\mlflow
\mlflow\1
\mlflow\1\{guid}\artifacts
\mlflow\1\{guid}\metrics
\mlflow\1\{guid}\params
\mlflow\1\{guid}\meta.yaml
\1\{guid}\checkpoints

i.e. the checkpoints are in the wrong location, they should be in the \mlflow folder.

Perhaps this is an mlflow rather than pytorch-lightning issue?

I'm using pytorch-lightning 0.8.5 on macos running in python 3.7.6

[Borda]

@david-waterworth mind try the latest 0.9rc12?

[awaelchli]

It was fixed here: #2502
The checkpoints subfolder will go here: mlflow\1{guid}\checkpoints, is that what you want @david-waterworth ?

[david-waterworth]

Thanks @awaelchli yes that's what I want - thanks!