Add declarative trainer

[Minyus]

Fixes # 912

Description:
Higher API for training will improve usability of Ignite.

Check list:

• New tests are added (if a new feature is added)
• [x ] New doc strings: description and/or example code are in RST format
• Documentation is updated (if required)

[sdesrozis]

@Minyus Thank you very much for this PR !!

I make you my feedback on your proposal quickly :)

[vfdev-5]

@Minyus thanks a lot for the PR ! I really appreciate your contribution, however it would be great to discuss more about the API. We already stated some general remarks on the API in the related issue #912

In my opinion, class NetworkTrain has very specific integrations with MLflow which can be a limitation for some users. Another point is about exposing trainer in the higher API in case of adding more user-specific handlers.

please, let's discuss more about the API in #912.

[vfdev-5]

@Minyus can we work out this at first, please ?

class TimeLimit:
    def __init__(self, limit_sec=3600):
        self.limit_sec = limit_sec
        self.start_time = time.time()

    def __call__(self, engine):
        elapsed_time = time.time() - self.start_time
        if elapsed_time > self.limit_sec:
            log.warning("Reached the time limit: {} sec. Stop training".format(self.limit_sec))
            engine.terminate()

This looks good enough. Maybe, in addition, we can set Checkpoint or ModelCheckpoint to save the work before terminating ?

Idea is to put it directly into ignite/handlers/time_limit.py

[Minyus]

@vfdev-5
OK, I'll move TimeLimit to ignite/handlers/time_limit.py.
Let's add saving before terminating to ToDo list as it needs further consideration (e.g. what if model checkpoint path is not specified, etc.)

[vfdev-5]

@vfdev-5
OK, I'll move TimeLimit to ignite/handlers/time_limit.py.
Let's add saving before terminating to ToDo list as it needs further consideration (e.g. what if model checkpoint path is not specified, etc.)

@Minyus thanks, but it would be much better to split this PR into several ones. One PR by feature.
So, please, send the code of TimeLimit into a new PR. It would be nice to follow https://github.com/pytorch/ignite/blob/master/CONTRIBUTING.md
Thanks

[Minyus]

As discussed in Slack, this PR is suspended.

Meanwhile, please feel free to send PR (including adding documentation, test, example, etc.) to PipelineX so that I can send a complete PR to Ignite later on.

higher API for training
https://github.com/Minyus/pipelinex/blob/master/src/pipelinex/ops/ignite/declaratives/declarative_trainer.py

MNIST example to use higher API
https://github.com/Minyus/pipelinex/blob/master/examples/mnist/mnist_with_declarative_trainer.py

TimeLimit handler
https://github.com/Minyus/pipelinex/blob/master/src/pipelinex/ops/ignite/handlers/time_limit.py

Cohen Kappa Score metric
https://github.com/Minyus/pipelinex/blob/master/src/pipelinex/ops/ignite/metrics/cohen_kappa_score.py

Please also feel free to copy any code in PipelineX to use in PyTorch Ignite.
I'm happy to see Ignite to be enhanced in any way.

[sdesrozis]

Thank you again and see you soon !!