* Fix serious bug in nesterov momentum optimizer.

* wip add se-resnext test for 2x2 and parallel exe

* testing

* finish dist unitest 2x2

* update python interp

* fix gpu number

* make this test can run in parallel

* Fix Reduce and Gather

* Fix unit test

fix AUC op
optimize it's test

* remove block api

* remove clone_variable

* hide block inner apis

* update

* fix tests

* fix auc

* update

* update

* fix compile

* fix param name

* add doc string

* fix test

* Add learning rate decay test

* fix test name

* doesn't share @LR_DECAY_COUNTER@

* refine multi-thread CPU Parallel exe

* refine multi thread CPU Parallel exe

* Refine CPU version for ParallelExecutor

* add share_parameter_between_cards_

* Fix ParallelExecutor bug

* Fix unit test

* Fix parameter opt balance

* Fix with opti (param->grad)

* Add grad to op var

* Remove shard_param_between_cards

Fix the HierarchicalSigmoidGradOpKernel and refine the codes. Now hsigmoid_op is same with V2 implementation and can pass gradient check.

* change im2sequence for ctc batch inference

* Update im2sequence_op.cc

* change im2sequence for ctc batch inference

* update

* change PR by comment

* fix ocr test error

* fix test_im2sequence

* modify the old name to standard name

* fix test_layers failed

* Add flag to set prior_box output order.

* Add a fake_quantize_op, which quantize an input tensor to a tensor with lower bits.

* Add region proposal network (RPN) target assign operator and Python API for Faster-RCNN.

* Fix batch norm when only 1 elements in normzalize dimension during training.