* fix predict

* fix export model

* fix doc

* Fix the timing of hubserving

* Fix "tar" to "pdparams" to adapt to dygraph

* Update the download link of Paddle Inference Library

* support fp16 training

* Use compiled training program

* Change timing ips.

* Use dali

* add pure fp16 training

* fix a bug, which will not use fuse pass using pure fp16 training.

* modify code as review

* modify loss, so that it will use different loss when using pure fp16 training.

* remove some fluid API

* add static optimizer.

* fix eval script

* fix dali shell

* fix pred

* fix resnest

* fix hrnet se

* fix se for export

* add shells, test=develop

* add support for multi cards eva

* fix local rank
* fix export model

* add static running in dygraph

add mkldnn config

* unify model and model_name

* rm unsed archs fix bug of se_resnext

* fix distillation model for distributed.launch

* fix export model for googlenet

add hub serving code and doc

* add densent init

* fix export model

fix python and cpp inference

fix feature map visualization

fix log benchmark

1. fix dist training 
2. fix cpp infer to support dir inference