Add transformer beamsearch decoder

Apply exponential decay on learning rate

* Fix bug in data anchor sampling.
* Add argument for memory optimization.
* Change arguments in detection_output.

Merge branch 'develop' of https://github.com/PaddlePaddle/models into add-transformer-BeamsearchDecoder-clean

* Add data anchor sampling in PyramidBox paper.

* update README_cn.md

* Refine code for English dataset.
1. Remove a pooling layer.
2. Change classes_num to 94.
3. Modify some arguments in ctc_train.py
4. Add learning rate decay policy.

* Fix readme.

* Fix README.

* Remove consine decay.

* Remove eval.sh

* Add icnet.

* Fix code.

* Fix utils.py.

Refine Pyramidbox infer and Pyramidbox configure

Refine Transformer for wmt14_en-de

* Refine documentation and code.

Update md5sum of aishell data

* Change the transposed conv2d initializer.

* small fix.

* Add more model configure, such as AlexNet, VGG, GoogleNet, DPN, ResNet etc.
* Add download_imagenet2012.sh
* Update doc.
* Add Chinese doc.

Adapt to fill_batch_size_like and make outputs between fast_infer and the original python infer alignment in Transformer

* DQN Atari based on RLLab

* DQN Atari based on RLLab

* DQN Atari based on RLLab

* refactor code without RLLab framework

* add module of saving and loading policy model

* refactor code structure, add DoubleDQN and DuelingDQN modules

* add fluid argmax, flatten utils

* update README.md

* udpdate replay memory

* udpdate replay memory

* update readme, code clean

* clean code and fix codestyle

* fix codestyle

* update README.md

* revisions|history->context, randint->random

* revisions| add comment for max-pooling operation in atari

* update wiki

* refine wiki

* train.py default args

* update wiki cn

* update wiki

fix tanh problem

* Add head bbox for Pyramid-Box model.

* Fix bug in transform_labels and satisfy_sample_constraint function.

fix inference bug

* Add object_coverage constraint & fix bbox_labels data layout bug

* code clean