Fix grad debug event

[typhoonzero]

Fix #3714

[lcy-seso]

LGTM.

[lcy-seso]

Just some questions,

1. Dose this PR means if I want to get parameter gradients, I should handle the EndForwardBackward event ?

2. Why I can still get gradients in CPU mode previously if just handle the EndIteration event, is there any difference in CPU and GPU training?

[typhoonzero]

Dose this PR means if I want to get parameter gradients, I should handle the EndForwardBackward event ?

Yes, you must. If not all gradients will be zero.

Why I can still get gradients in CPU mode previously if just handle the EndIteration event, is there any difference in CPU and GPU training?

Please check out https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/trainer/ThreadParameterUpdater.cpp#L201, in v2 API training, default creates SgdThreadUpdater, and the default parameter update in GPU put all gradients to zero immediately after update

[lcy-seso]

I see, thank you! I think I should also update demos in the model repo after this PR.