* deepfm dygraph

* deepfm dygraph fix readme

* adapt dygraph dev api and fix readme

* fix eval and readme

* deepfm dygraph add infer.py, format train output

* fix bug of last commit

* add more log in train and infer. remove is_sparse

* update build_once for cycle_gan
test=develop

* add input_channel arg
 for cycle_gan
test=develop

test=develop

test=develop

* modify ce for deeplabv3+

modify ce for auto_dialogue_evaluation

add ce for human_pose_estimation

* modify gru4rec ce

* support out json in eval and add exclude fc params in finetune

* fix cpu bug

* add prediction file

* modify prepocess

* add assertion

* modify senta dygraph

* modify details

* update ptb model to use auto generated op functions, test=develop

* refine model code, test=develop

* refine model code, test=develop

* update ptb model, test=develop

* add ce for PaddleDetection

1. use set_value() to set weight.
2. use weight in PositionEmbedding instead of protected member `_w`.
3. fix data layout for Embedding, the tailing sized-1 dimension is removed, update documentation, 
4. fix tensorboard tags, spaces are replaced with `_`.

* fix paddlerec dnn loss nan

* change hash with mmh3.hash in dnn

* fix get cuda device bug when only use cpu

* fix typo

* fix step time in vae

* fix step time in vae