* Add py-reader and parallel-executor support in Transformer inference

* Add statick k, v cache for encoder output in Transformer inference

* Replace the cache from compute_qkv with cahce from split_heads in Transformer inference

* Fuse k, q, v projection in Transformer

* Revert the fused k, q, v projection in Transformer to be compatible with saved models

* Use gather_op to replace sequence_expand_op in Transformer inference

* Add fluid_transformer.md

* Refine README for released models and data in Transformer

* Refine README for released models and data in Transformer

fix .sh params

add video classification models in PaddleCV

refine image classification train script

Merge branch 'video_classification' of https://github.com/SunGaofeng/models into video_classification

Merge branch 'video_classification' of https://github.com/SunGaofeng/models into video_classification