learning_rate adjustment for run_multi_process, run_single_process

First of all, I really appreciate the clean design and abstraction of dynet/mp. When I read the code of ILearner, SufficientStats, Datum, run_single_process and run_multi_process, I finally find that it is time to stop writing tedious training function to loop over training dataset and report every N updates, test on dev dateset every epoch, and keep track of a loss accumulator or sth like that.

When using run_single_process or run_multi_process, just write a function for a single instance is enough. It is really cool.

But the problem is how to update the learning rate of the trainer when using these high-level abstractions. The run_child function does not include learning rate update. Is there a clean way to add code for that in the same level of abstraction?

Since update_epoch() has been deprecated, and there is no way to pass something like a learning rate scheduler to run_multi_process and run_single_process.

Start by reading ILearner, SufficientStats, Datum, run_child, run_single_process, and run_multi_process, then compare their training flow with the deprecated update_epoch(). Determine how a learning-rate scheduler or equivalent update hook could be passed through both high-level runners. Done means learning-rate updates can be configured without manually rewriting the training loop.