* Error out if fp16 is enabled and affine channel is used

* Add more documentation on mixed precision training

* add multi-scale test

This reverts commit b591cc133ca04dc1ed50631c8d24031bb97db18b.

* Add face detection configs and main code.
* Integration multi-scale evaluation.

* unify reader to dataloader

* add check_version

* refine build strategy for the model optimize test=develop

* fix according develop test=develop

* Place mixed precision inside PaddleDetection

roll back to the monkey patch version as a temporary measure, before it is
merged into paddle

* Add command flag for `loss_scale`

* Fix a stupid indentation error

optimizer should be in the mixed precision context

* Initial FP16 training

* Add mixed precision training to rest of the detection models

* Revert "Add support for mixed precision training (#3406)"

This reverts commit 3a2c106271885071db7c0d85587540a8f83c24db.

* Bug fixes and some tweaks

* Support finetune by custom dataset

* add finetune args

* add load finetune

* reconstruct load

* add transfer learning doc

* add fruit demo

* add quick start

* add data preprocessing FAQ

* Disable optimizer fusion for RCNN models

waiting for upstream fix

* Disable completely

* Add support for mixed precision training

* Disable `fuse_all_reduce_ops` when training in mixed precision

waiting for upstream fix

* Initial support for distributed training

* Housekeep on rank 0 only

* Conform to models convention

In PaddleDetection,  making use of tb-paddle to record data, so as to display in TensorBoard.  (#3295)

* use tb_paddle to log data
including:

* loss
* mAP
* image

* modified:   requirements.txt
* modify the missing argument --use_tb
* modify GETTING_STARTED.md and docs/GETTING_STARTED_cn.md

add the argument of tb-paddle in docs.
* optimize the display name of image displaying.

* Add benchmark (fps) in MODEL_ZOO doc

* fix add map_type

# Add VGG-SSD on VOC and COCO dataset.
# Add config and model zoo. 
# Refine bbox2out and draw_bbox:
    - Add bbox de-normalize in bbox2out function.
    - Remove bbox de-normalize in draw_bbox.

* add save best model & fix dcn configs

* fix yolov3 voc eval in train

* fix comment

* fix easy problem

* add voc_eval and yolo_darknet_voc

* add yolov3_darknet_voc in MODEL_ZOO

* fix default im_size

* fix MODEL_ZOO note

* fix is_bbox_normalized

* extract map to map_utils.py

* update yolov3_dorknet_voc mixup

* add yolov3_r34_voc

* add yolov3_mobilenet_v1_voc

* fix drop empty in VAL mode

* use cfg.num_classes

* assert metric valid

* enable difficulty can be None

* add comment for bbox_eval

* num_classes in retinanet

* add log iter argument

* ppdet log time smooth

* ppdet log eta