Able to print gradients in event_handler

[typhoonzero]

Able to print gradients in event handle using v2 API to train.

Related: #3040
Fixes: #3211

def event_handler(event):
        if isinstance(event, paddle.event.EndIteration):
            if event.batch_id % 10 == 0:
                for p in parameters:
                    print "parameters:", parameters.get(p)
                    grad = parameters.get_grad(p)
                    print "gradients:", grad
                    print "gradients max abs:" max(grad.min(), grad.max(), key=abs)
            if event.batch_id % 100 == 0:
                print "Pass %d, Batch %d, Cost %f" % (
                    event.pass_id, event.batch_id, event.cost)

[Yancey1989]

Just a little confusing, why we move the batch_evaluator to the event_handler?

[typhoonzero]

Need to call event_handler before finishBatch operations so we can get inner status before clearing them.

[Yancey1989]

Got it, thanks!

[jacquesqiao]

rename to get_param?

[typhoonzero]

__getitem__ will called when doing param[k], it's an operator reload. Need to keep this the same as before.

[lcy-seso]

even though it copies the parameter from the c++ side, very helpful for debugging the training process and tuning the model, thank you very much.