* add new load dygraph func

* update load_pretrain_params

* update load_dygrah_params

* Update save_load.py

* Update train.py

* Update save_load.py

* return {} when path is None

* return {} when path is None

* 识别模型导出

* 识别模型inference

when you trian the det model , there is not the var: loss_type in the Global, so  train_alg_type == 'rec' is necessary