LoggerCollection breaks PyTorch profiler

[gahdritz]

🐛 Bug

If the Trainer's profiler parameter is set to "pytorch" and the Trainer's logger is an instance of LoggerCollection, the profiler fails to write to a local file (with a warning).

The path for said file is derived from this property of the Trainer, which in turn derives from the save_dir of the Trainer's logger whenever the logger isn't a stock TensorBoard logger. The save_dir property of a LoggerCollection, which is what you get when the trainer is provided with more than one logger, is set to None by default, apparently deliberately. I'm not sure what the rationale behind that decision was, so I'm hesitant to make a PR myself, but it seems logical to me that LoggerCollection should have a log_path property defined analogously to that of a TensorBoardLogger, and that the property of the Trainer shouldn't fall back to the save_dir property automatically when its logger isn't an instance of TensorBoardLogger.

Reproduction

Here's a BoringModel Colab link. The following warning is generated during training when the Trainer's logger is wrapped in a list:

/usr/local/lib/python3.7/dist-packages/pytorch_lightning/profiler/pytorch.py:441: UserWarning: The PyTorchProfiler failed to export trace as `dirpath` is None
  rank_zero_warn("The PyTorchProfiler failed to export trace as `dirpath` is None")

Expected behavior

The PyTorch profiler should always have access to a default logging path, even when the Trainer's logger is a LoggerCollection. Ideally, it would follow the same "version" convention obeyed when the logger is a TensorBoardLogger.

Environment

* CUDA:
	- GPU:
		- Tesla T4
	- available:         True
	- version:           10.2
* Packages:
	- numpy:             1.19.5
	- pyTorch_debug:     False
	- pyTorch_version:   1.9.0+cu102
	- pytorch-lightning: 1.3.7post0
	- tqdm:              4.41.1
* System:
	- OS:                Linux
	- architecture:
		- 64bit
	- processor:         x86_64
	- python:            3.7.10
	- version:           #1 SMP Sat Jun 5 09:50:34 PDT 2021

[tchaton]

Dear @gahdritz,

Thanks for reporting this issue. Would you be interested in making a fix PR to Lightning ?
I believe it shouldn't be too hard and a good way to learn more :)

Best,
T.C

[gahdritz]

Yeah I can take a crack at it

[gahdritz]

@tchaton the PR is done. Any feedback is welcome.