fix tsn pretrain load ResNet50

fix bug

* refine README in LRC

add resume weights in train.py

* add LRC Local Rademachar Complexity Regularization

* fix bug in eval faster & refine code

* Delete file
* Rever the raw file

This reverts commit 52fe0fe8.

Add build_raw_data.py

* Adjust the structure of whole repo

* Correct the readme for speec

* Update submodules LAC, Senta and SimNet

* fix reader bug in RCNN

fix config of AttentionCluster/LSTM

* 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