Add `lr_multi_list` param for SSLD pretrained models' usage.

* activate scale_x_y

* update cspdarknet pretrained model

revert `reshape` changes

* optimize python path

* add yolov4

* Revert back to `fluid.clip` for gradient clipping

current API is in flux, wait for it to stabilize

* Please CI

still buggy

* #add dcn on  FCOS_head and backbone

* code for cornernet

* add doc

* refine custom op

* refine code from review comments

* refine doc

* update doc

* update code

* add cornernet_squeeze_mixup_cosine config

* update code

* update code

* add moile-side rcnn server model

* 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 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%

fix multi-scale def

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

* Replace `switch` ops with `cond`

* Use func instead of lambda

* exit1 in travis failed

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