More user-friendly Save/Load API of Fluid

[QiJune]

We now have three save interface in fluid:

• save_vars for saving a list of variables
• save_params for saving all parameters
• save_persistables for saving checkpoint

And save_vars interface is not exposed to users directly since users can not access a variable created inside a layer. We can use save_persistables to do checkpoint when training. And we expect save_params can work in saving model for inference.

But save_params and save_persistables are not enough to cover all cases. For example, if we want to save a model for inference network with BN layer, global mean variable is needed, whereas momentum variable is not needed. Both save_params and save_persistables are not satisfied.

We need to design a more user-friendly save/load API.

[reyoung]

Just adding a reference.

Tensorflow uses global sets of tensors, which are named collections, to identify the kind of a tensor. The collections could be these.
Users can save any parameters in any collection.