* add ppo

* fix bugs

* yapf

* torch benchmark policy gradient

* refine comments and use native api

* add pytorch a2c

* add set/get_weights test & copyright

* yapf....

* Update model_base_test_torch.py

* update

* Delete banma.py

* Update model_base_test_torch.py

* update

* Update model.py

* update torch tests

* Update model_base_test_torch.py

* git commit -m torchdqn

* yapf

* fix bugs

* fix bugs

* fix bugs

* yapf

* remove fstring format

* torch_test yapf

* yapf

* Add torch in unittest.requirements

* update torch_unittest

* Torch and FLUID conflict problem in __init__.py

* Unittest fail for torch when both torch and fluid exists.

* cluster_test fail in the unittest, add timeout seconds.

* Torch backend for PARL

* add sleep time for unit test send_job_test.py

* Unit test for send_job_test.py

* use multiple try for unit test

* Fix compatibility for python2.7.

* fix send_job_test.py bugs

* check file exist before send_job_test.py

* Modify send_job_test.py

* add IMPALA algorithm and some common utils

* update README.md

* refactor files structure of impala algorithm; seperate numpy utils from utils

* add hyper parameter scheduler module; add entropy and lr scheduler in impala

* clip reward in atari wrapper instead of learner side; fix codestyle

* add benchmark result of impala; refine code of impala example; add obs_format in atari_wrappers

* Update README.md

* sync paras in program, fix deepcopy bug, python3 compatibility

* refactor code, add plutil directory, clean import order

* remove old comment

* refine comment

* fix codestyle

* cache sync program, add gputils module, refine model_base unittest

* fix codestyle

* refine sync params cache

* add fetch_value module

* redesign basic class in PARL

* code style fixed

* update yaml's version

* update yaml's version & update code to fix style problem

* add debug message for  function

* delete test code

* rename function: has_fun -> has_func

* prelimary implementations of ComputationTask, Algorithm and Model classes

* remove "model_func" from the args of an algorithm

* a clean clone() function for Algorithm and Model

* add use_next_value as a input to learn()

* further re-structure

* added Feedforward and RLAlgorithm classes

* maxid -> argmax

* discrete_distribution -> category_distribution

* category -> categorical

* revisions