[python] Use first_metric_only flag for early_stopping function.

[matsuken92]

In order to apply early stopping with only first metric, applying first_metric_only flag for early_stopping function.

[matsuken92]

The following error may be CI environment problem on appveyor. TravisCI and Azure DevOps have same issue. How to deal with it to solve...?

https://ci.appveyor.com/project/guolinke/lightgbm/builds/23063424/job/7unpekbvoh1gwt3r

jupyter nbconvert --ExecutePreprocessor.timeout=180 --to notebook --execute --inplace *.ipynb
Traceback (most recent call last):
  File "C:\Miniconda36-x64\envs\test-env\Scripts\jupyter-nbconvert-script.py", line 6, in <module>
    from nbconvert.nbconvertapp import main
  File "C:\Miniconda36-x64\envs\test-env\lib\site-packages\nbconvert\__init__.py", line 7, in <module>
    from . import postprocessors
  File "C:\Miniconda36-x64\envs\test-env\lib\site-packages\nbconvert\postprocessors\__init__.py", line 5, in <module>
    from .serve import ServePostProcessor
  File "C:\Miniconda36-x64\envs\test-env\lib\site-packages\nbconvert\postprocessors\serve.py", line 19, in <module>
    class ProxyHandler(web.RequestHandler):
  File "C:\Miniconda36-x64\envs\test-env\lib\site-packages\nbconvert\postprocessors\serve.py", line 21, in ProxyHandler
    @web.asynchronous
AttributeError: module 'tornado.web' has no attribute 'asynchronous'
Command exited with code 1

[guolinke]

@StrikerRUS can you help to review this? Thank you!

[StrikerRUS]

@matsuken92 Thanks for your contribution! Can you please write a test for the enhancement you are proposing? I think it'll be rather straightforward with self-defined dummy metrics.

[msftclas]

All CLA requirements met.

[StrikerRUS]

I went ahead and added simple test and updated docs with info about the new feature.

[guolinke]

one question:
when there are multi validation sets, the training will stop when any of them met early stopping? or just the first validation set?

[StrikerRUS]

I think that the behavior with multiple valid sets should remain unchanged:

Early stopping requires at least one set in valid_sets. If there is more than one, it will use all of them except the training data:

[guolinke]

@StrikerRUS @matsuken92 I think this implementation is only for the first metric of first validation set.

[StrikerRUS]

Then it should be addressed in #2127 to preserve documented behavior, I think.