Add CSV logging to LighterLogger

[ibro45]

Description

Ideally, match all the loggers that PyTorch Lightning provides.

• TensorBoard
• Wandb
• Comet
• MLFlow
• Neptune
• CSV logging (@kbressem I remember you have something like this in your trainlib, suggestions are welcome)

[kbressem]

Yes, I log train loss on a step level and metric on an epoch level. This allows to create nice graphs, similar to tensorboard. Also, I save individual metrics per file, but currently they are overwritten each epoch, this could be changed. So it is not really logging but more continuously saving results. Still very helpful for me.

Regarding how I do it. I just pull each metric from the engines state and dump it into a file, separating each value with a comma. I do not use pandas or any external classes/functions from monai or ignite. So nothing fancy.

CSV looks like this:

filename	class 1	class 2	class n
/path/to/dicom1	0.87123	0.74623	0.12341
/path/to/dicom2	0.45425	0.84234	0.77412

I could imagine that logging every individual metric on an epoch level would also be helpful, to identify cases where the model has problems.

[ibro45]

@surajpaib and I discussed - we will not add support for other types of logging, we will have Tensorboard, W&B, and CSV logging support. Still need to implement the CSV one.

[ibro45]

#93 - LighterLogger removed, relying on PL loggers.

Should use https://lightning.ai/docs/pytorch/stable/extensions/generated/lightning.pytorch.loggers.CSVLogger.html now. Haven't tried it ever. Have you guys?