add deform_conv (#52)

# add dcn in fcos_head.py(#95)

* add iou aware module
* iouaware model zoo
* fix copyright

* add mobilenet v3 configs

* fix some comment

* fix memsize

* fcos initialization

* delete the codes not used

* pre-commit the committed code

* delete the unused function

* modify the capacity in loader from 64 to 16

* add mask operations in RandomExpand and RandomCrop

* remove all [] in cropped polys

* is_poly -> _is_poly()

* change the judgment on whether there are polygons

* abstract is_poly to op_helper.py

* add is_mask_expand and is_mask_crop in RandomExpand and RamdomCrop

* add mask operations in RandomExpand and RandomCrop

* remove all [] in cropped polys

* is_poly -> _is_poly()

* change the judgment on whether there are polygons

* abstract is_poly to op_helper.py

* add devices_num on reader

* refine capacity value

* refine bufsize

* add yolov3_mobilenet_v3

* add server side rcnn model, 62 fps with coco mAP 41.6% and 20fps with coco mAP 47.8%

* export config file in export_model.py

* add custom label_list & dump_box

* add infer_cfg.yml & doc

* export label_list to inference config

* update PaddleTensor to ZeroCopyTensor

Add cosine decay lr strategy.

fix multi-scale def

* add autoaugment, which is validated on faster-rcnn

* add multi-scale testing function in cascade rcnn clsaware architecture

* refine transfer learn

* refine transfer learning doc

* Replace `switch` ops with `cond`

* Use func instead of lambda

* Update checkpoint to use paddle 1.7 API

* Add new checkpoint support to `load_and_fusebn`

* Make yapf happy

* enable data unittest

* exit1 in travis failed

* fix all documentation

* fix docs conf.py

* fix some error

* fix readthedocs

add libra rcnn, including
  * libra loss
  * libra fpn(bfp)
  * libra sampling