Logger and log dir names

1

A related issue with logger and log dir namings. Let’s say I want the following loggers:

# A
pl_loggers.TensorBoardLogger(save_dir=os.getcwd(),
                             version=None,
                             name='lightning_logs'),
# B
pl_loggers.WandbLogger(save_dir=os.getcwd()),

Logs saved with Trainer(…, logger=[A]) :

        lightning_logs
        ├── 0
        │   └── checkpoints
        │       └── epoch=4.ckpt
        └── version_0
            ├── events.out.tfevents.1601074580.ip-172-31-95-173.103868.0
            └── hparams.yaml

Logs saved with Trainer(…, logger=[A,B]) :

lightning_logs
└── version_0
    ├── events.out.tfevents.1601071061.ip-172-31-95-173.63314.0
    └── hparams.yaml
lightning_logs_<project_dir_name>
└── 0_210yiyz0
    └── checkpoints
        └── epoch=24.ckpt

Logs saved with Trainer(…, logger=[B,A]) :

<project_dir_name>_lightning_logs
├── 3n3bfyoa_0
│   └── checkpoints
│       └── epoch=29.ckpt
lightning_logs
├── version_0
│   ├── events.out.tfevents.1601073059.ip-172-31-95-173.86365.0
│   └── hparams.yaml

Finally, logs saved without explicitly passing logger:

lightning_logs
├── version_0
│   ├── checkpoints
│   │   └── epoch=39.ckpt
│   ├── events.out.tfevents.1601074683.ip-172-31-95-173.105180.0
│   └── hparams.yaml

The last output structure/naming is desired. However, that seems to conflict with the requirement of adding logger B the w&b logger.

2

Hi
Agree, this is something that I also find annoying.
I tried to make this work awhile ago, but I didn’t get this to work for WandB.
It does not seem to be possible to change the default logging folder for wandb.
Maybe it changed in the latest version, I will have a look.