* adam sparse support

* fix gpu build error

* fix ci

* fix ci

* fix adagrad sparse update bug

* fix gpu build error

* add decayed adagrad python code

* fix typo and order

* small fix

The size of the bias parameter should be the number of filters.

- Removed all main_program and startup_program in the demo.
- Using paddle.default_main_program() hides the implementation detail (e.g., using g_main_program) from the user, we can change the implementation in the future much easier.

* Adding greater than and less than equal ops to compare op
* Changing the name of the less_than_equal and greater_than_equal op
* Also changing the name of the functors

* init commit

* change some dir name

* fix lod_tensor_array

* init test beam search decode op

* add test_beam_search_decode_op

* Conditional Block Forward

* Assign Operator.

Out=X, when type in [LoDTensor/SelectedRows/LoDTensorArray]

* Stash

* Add Scope::Rename

it is useful in gradient phase of an operator with block

* ConditionalBlock Grad Done

* Add comments

* yapf format code

* add split lod tensor operator

* add more test cast

* clean code

* add merge lod tensor operator

* fix bug

* clean code

* add grad operator

* make mask support GPU

* add comments

* Assign Operator.

Out=X, when type in [LoDTensor/SelectedRows/LoDTensorArray]

* Follow comments

* add lstm layer

* set hidden shape

* rename input parameter

* add dynamic lstm

* refine dynamic lstm layer

* change parameter using XavierInitializer by default

* refine dynamic lstm layer

* create learning rate variable for every parameter

* fix ci

* set parameter lr relatively to global lr

* "fix ci failed"

* "comment out seq_concate op to unblock PRs"