fix duplicate console logging bug / properly configure logging

[awaelchli]

What does this PR do?

Fixes #4621

Current Behavior

The following code snippet reproduces the duplicate logs shown in the console:

import pytorch_lightning as pl
import logging
logging.info("I should not get logged, because logging is not configured!")
pl.seed_everything(1234)  # this gets logged twice

# console output:
# Global seed set to 1234
# INFO:lightning:Global seed set to 1234

Expected Behavior

We expect the following behaviors:

Scenario 1: User did not configure logging, we provide a default StreamHandler to stderr that does not propagate

import pytorch_lightning as pl
pl.seed_everything(1234)
# console output:
# Global seed set to 1234

Scenario 2: User configures logging, expects Lightning logs to propagate

# logging configuration
import logging
logging.basicConfig(level=logging.DEBUG)  # custom formatting etc.
import pytorch_lightning as pl
logging.info("hello!")
pl.seed_everything(1234)
# console output:
# INFO:root:hello!
# INFO:pytorch_lightning.utilities.seed:Global seed set to 1234

This PR will properly configure python logging for the library. The messages get all propagated in the module hierarchy as it is recommended by python logging best practice. This gives the user the most flexibility, they can now configure logging on a module level, see the updated docs for an example.

Correct Logging Usage

In the future, if you want to use logging as part of the Lightning library, please do:

import logging
logger = logging.getLogger(__name__)    # <-- good!

from pytorch_lightning import _logger   # <-- bad!
logger =logging.getLogger("lightning")  # <-- bad!

NOTE: This PR has many files changed, however the changes are small and only in the header the files. Most important changes can be found in these files:

pytorch_lightning/__init__.py 
docs/source/logging.rst
tests/test_profiler.py

works fine in colab too
https://colab.research.google.com/drive/1pgKFHcPkSxCRydp84PPunHQibNB-fZLq?usp=sharing

[codecov]

Codecov Report

Merging #5509 (6a500f9) into master (f094e10) will decrease coverage by 0%.
The diff coverage is 96%.

@@          Coverage Diff           @@
##           master   #5509   +/-   ##
======================================
- Coverage      93%     93%   -0%     
======================================
  Files         135     135           
  Lines       10005   10023   +18     
======================================
+ Hits         9340    9356   +16     
- Misses        665     667    +2     

[Borda]

43 files changed is a bit scary :D

[awaelchli]

@Borda the change in each file is minor, just in the header of the file I replace

from pytorch_lightning import _logger

with

import logging
_logger = logging.getLogger(__name__)

I see no other way. To do the logging properly we want to respect the module hierarchy and this requires changing all occurences of logging.

[carmocca]

See comments, also, all these logs will not be filtered to appear just on rank zero, right?

[justusschock]

@carmocca They weren't filtered before and they shouldn't be, since these are logs intended to be displayed on each node for investigation purposes.

[Borda]

@carmocca mind re-review?

[Borda]

lets be consistent and use the same name as you defined in __init__:

_logger = python_logging.getLogger(__name__)