Do not configure python logging

[rmrao]

🐛 Bug

pytorch-lightning right now configures the python logging module (here). This is generally not recommended when writing a library as it makes it difficult for users to modify logging format (see python docs, Stack Overflow post). I would suggest deleting the configuration line.

[hadim]

I agree. Logging configuration should be disabled by default. We could still set up the logging configuration manually just after importing PL.

[Borda]

we had the discussion here #1267 (comment) @jeremyjordan ^^

[rmrao]

Hmm... ok. In that case, could I potentially suggest that the call to basicConfig be placed in Trainer's __init__? Otherwise you require the user to configure logging before importing pytorch lightning. There are a couple problems with this:

1. imports should generally be placed at the top of a file. Configuring logging before importing pytorch-lightning means users have to violate PEP8 at some point.
2. It's even more difficult when using DDP, since it's not possible to configure the logging for a DDP process before importing pytorch lightning.

[jeremyjordan]

We could still set up the logging configuration manually just after importing PL.

@hadim isn't that what we do now? we configure logging when pytorch-lightning is imported

could I potentially suggest that the call to basicConfig be placed in Trainer's init?

@rmrao i'm not sure this would help you since we import Trainer in the project's root __init__.py

however, if you configure logging as you wish in your project's root __init__.py, by the time lightning is imported the root handler will already be configured and basicConfig will not run.

tagging @williamFalcon since we had this discussion here already. would prefer not to keep going back and forth on how we configure logging for the project.

[Borda]

We could still set up the logging configuration manually just after importing PL.

@hadim isn't that what we do now? we configure logging when pytorch-lightning is imported

now it sets the level for all logging not just for this lightning logger, that is what I was talking before...

[jeremyjordan]

just doing something like

logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)

doesn't actually set up a handler to send the log messages.

however, we could just simply add a line to configure a stream handler

logger = logging.getLogger(__name__)
logger.addHandler(logging.StreamHandler())
logger.setLevel(logging.INFO)

[rmrao]

@jeremyjordan my suggestion is actually to put basicConfig inside of the Trainer's __init__ function. This would only call basicConfig when an instance of Trainer is created (not when it's imported). You could even include the log level as an argument to Trainer.

[jeremyjordan]

@rmrao ah i see, i misunderstood your suggestion. i think that's a fine solution

[fjhheras]

I was having problems because of this. It would duplicate my log messages.

I understand this will be solved soon, but in case anyone has the same problem, this is how I circumvented it. When first defining the root logger, I do:

    logger = logging.getLogger()
    # if there is already a handler, remove it.
    if logger.handlers:
        logger.handlers.pop()

(logging is the python logging package, not pl's)

[FrancescoSaverioZuppichini]

same here, how do I use a logger from logging?