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.

* make transpiler test reliable

* add more

* follow comments

* Add anchor generator operator for Faster-RCNN.
* Add unittest testing.
* Add Python API.

* memory init

* add env

* refine anounce

* Add check for Nan

* Debug

* Add env for cc_test

* Add env for py_test and nv_test

* Remove py_test env

* Add env for py_test

* serial test_recognize_digits

* Test FLAGS_init_allocated_mem function for unit test

* Init allocated mem for op unit test

* Add env for all unit test

This reverts commit bc28cf61.

* Fix mean iou op.

* Fix crop layer test.

* Fix unitest

* fix unitest.

* Add fill_zeros_array op. This op is used for zero-filling an LoDTensorArray.

* merge fill_zeros_array_op with fill_zeros_like_op

* add unit_test for fill_zeros_like for array

* bnorm+relu fuse for mkldnn

* separate fuse_relu function

* bug fix

* proper while range in inference_transpiler

* description fix

* review fix

* review fix

* unit test for fwd batch norm+relu MKLDNN fuse

* optimize bipartite_match_op in large scale input

* Speed depthwise transposed conv2d.

* change return_numpy[PE] default value

* Remove convert to numpy in unit test