reduce travis-ci time of build_doc

[luotao1]

fix #3303
related #1854

1. build_doc don't depend on import py_paddle now, thus, remove import py_paddle in both conf.py.cn.in and conf.py.en.in
2. As build_doc depends on import paddle, we can use '@PROJ_ROOT@/python/paddle instead of installing paddle python package, which will reduce a lot of time.

build_doc in Dockerfile (build.sh) will be updated in next PR.

[wangkuiyi]

LGTM++

Thanks for this change!

Does it also fix #1854?

[wangkuiyi]

Should we use rm -rf output here? It seems that rmdir output works only when output is an empty directory.

[luotao1]

As the next line git clone $REPO output , we remain mkdir output here.

[wangkuiyi]

Wow. Does this mean that the build_doc Travis job doesn't build Paddle twice now?

[luotao1]

Yes, build_doc Travis Job doesn't build twice, and doesn't build the whole Paddle now.

[wangkuiyi]

Travis CI's build_doc jobs failed:

Generating sphinx documentation: html
Running Sphinx v1.5.6
making output directory...
Traceback (most recent call last):
  File "/home/travis/build/PaddlePaddle/Paddle/python/paddle/trainer/config_parser.py", line 85, in <module>
    from paddle.proto.DataConfig_pb2 import DataConfig
  File "/home/travis/build/PaddlePaddle/Paddle/python/paddle/proto/__init__.py", line 15, in <module>
    from paddle.proto.TrainerConfig_pb2 import OptimizationConfig, TrainerConfig
ImportError: No module named TrainerConfig_pb2
Configuration error:
The configuration file (or one of the modules it imports) called sys.exit()
make[3]: *** [doc/CMakeFiles/paddle_docs] Error 1
make[2]: *** [doc/CMakeFiles/paddle_docs.dir/all] Error 2
make[1]: *** [doc/CMakeFiles/paddle_docs.dir/rule] Error 2
make: *** [paddle_docs] Error 2

I tried to help fixing it, but no idea how to do it.

[Xreki]

There are some other places that import py_paddle, and I'm afraid that building of py_paddle cannot be removed. See as following:

$ grep -r py_paddle *
paddle/v2/data_feeder.py:from py_paddle import DataProviderConverter
paddle/v2/data_feeder.py:        :type argument: py_paddle.swig_paddle.Arguments
paddle/v2/inference.py:        import py_paddle.swig_paddle as api
paddle/v2/parameters.py:        import py_paddle.swig_paddle as api
paddle/v2/parameters.py:        import py_paddle.swig_paddle as api
paddle/v2/parameters.py:    import py_paddle.swig_paddle as api
paddle/v2/__init__.py:    import py_paddle.swig_paddle as api
paddle/v2/event.py:        import py_paddle.swig_paddle as api
paddle/v2/tests/test_parameters.py:    import py_paddle
paddle/v2/tests/test_parameters.py:    del py_paddle
paddle/v2/tests/test_data_feeder.py:import py_paddle.swig_paddle as api
paddle/v2/optimizer.py:        import py_paddle.swig_paddle as swig_api
paddle/v2/optimizer.py:        import py_paddle.swig_paddle as swig_api
paddle/v2/optimizer.py:        import py_paddle.swig_paddle as swig_api
paddle/v2/optimizer.py:        import py_paddle.swig_paddle as swig_api
paddle/v2/optimizer.py:        import py_paddle.swig_paddle as swig_api
paddle/v2/optimizer.py:    import py_paddle.swig_paddle as swig_api
paddle/v2/trainer.py:        import py_paddle.swig_paddle as api
paddle/v2/trainer.py:        import py_paddle.swig_paddle as api
paddle/v2/trainer.py:        import py_paddle.swig_paddle as api

[luotao1]

@wangkuiyi
The reason is that proto/ is not generated in python/paddle directory, i.e. we need /python/paddle/proto. I will fix it.

@Xreki
All these py_paddle are in functions, such as:

class Optimizer(object):
    def __init__(self, **kwargs):
        import py_paddle.swig_paddle as swig_api
        if 'batch_size' in kwargs:
            del kwargs['batch_size']  # not important for python library.

When building the document, it only read __doc__ of class or function, don't come into themselves.

[luotao1]

@Xreki py_paddle in data_feeder.py will be fixed in next PR. Thx very much for your kindly help.

[luotao1]

The travis time of build_doc from 35minutes to 5 minutes, and the total travis time is less than 10 minutes.