In the future, we'd like to encourage user to save
everything during training.
This allows us to
1. Do more flexible optimization passes
2. Re-train and fune-tune

fix py2 matplotlib deps

clean unused inference_optimize c++ implementation

* add unit test for nested drnn

* add nested dyn_rnn

* refine while_op

* fix bug

* Add truncated gaussian initializer.

* Fix unitest.

* Update API.spec

* Fix code style and fix bug.

* Fix code style.

* Small fix.

Modify some ops to save memory

Wait input when data transform

* Add python api for expand op.

* Fix unitest.

* Remove 'out' from arguments and fix code style.

* fix API.spec

* Fix API

* Fix unitest

* loosen the restriction of output_size in conv2d_transpose

* test and docs

* fix code style

* fix ci test error

* bug fix

* fix python3 issue

fix memory optimize

Speed up lod_tensor to array and array to lod_tensor

Hide RecordIO Reader

* hide program APIs

* fix merge error

* update

* refine seq_softmax_op

* fix seq_softmax

* use cub in seq_softmax

update install command of latest readme

Get sequence length in sequence_pad op & fix sequence_mask op

* fix dist table send hang problem

* revert sync_mode config

* fix async send table

Refine seq_concat

Refine  seq_pool

Speed up lod

Versioning doc

fix dso cmake typo

* Add sequence_expand_as_op

* follow comment

Fix transformer random failed

simplify and hide bcast_params