* Forget to finishTestPeriod in testOnePeriod.
* Fix #318

* Add benchmark for PaddlePaddle, tensorflow and caffe

* ConvProjection to reduce memory for goolenet

* Add unit test for ConvProjection.
1. unit test in test_LayerGrad.
2. compare the ConvPorjection and CudnnConvLayer, also compare the concat_layer+img_conv_layer and concat_layer_conv_projection.

* Reduce cudnn_conv memory and add benchmark document.
1. Use TmpMatrix as the workspace in cudnn_conv to reduce gpu memory. It reduce lots of memory.
2. Add benchmark document.
3. fix smallnet_mnist_cifar.py in paddle.

* Add job=time and refine cudnn_conv to reduce gpu memroy and speed up

* Refine cudnn_conv and shared biases operation in concat_layer and mixed_layer.

* follow comments

* follow comments

* Use unique_ptr to prevent memory leaks in CudnnConvLayer.

* Also fix #243 

* Python trainer API and demo

* Adding missing PaddleAPIPrivate.h

* Adding api_train.sh

* More comments

* Bump up patch version to 0b3

* Fix sparse training for trainer_count=1

For trainer_count=1, the gradient machine is NeuralNetwork, which does not create parameter buf for PARAMETER_GRADIENT for sparse update in Parameter::enableType. But gradient parameter buf is still used in SgdThreadUpdater.

* Minor update to comment

* add interface and unittest for nce layer

* follow comments

* Use C++ 11 atomic_flag in MacOS as spin lock
* Add unittest for it.

* follow comments to fix bugs

* refine sparse momentum api and unittest
* fix unittests bug

* reduce trainer count for unit test on MAC OSX

* Add `device` parameter to ExtraAttr in trainer_config_helpers.
* add unittest for it.

Local training with "sparse_update = True" parameter triggers kSgdSparseCpuTraining mode, fix bugs under it.

* remove unnecessary field set in ParameterConfig, Evaluators, etc

* min_pool_size would be infinite by default.
  * add unittest for min_pool_size
* Fix bug in can_over_batch_size
  * add unittest for can_over_batch_size
* Add DEFINE_PROVIDER_EX
* Add default value of should_shuffle
  * When training, the default value of should_shuffle is True.
  * When testing, the default value of should_shuffle is False.
  * User a set a provider should_shuffle or not by pass it to `@provider`
  * should_shuffle can handle a list of value, not just boolean
* Add input order mapping by using name
  * Add unittest
* Add check to check input format.
  * Default is close for speed reason.
  * User could stop train when check error, or continue train without
    this train sample.
* use deque instead of vector in generators pool, make erase
  generator faster.
* Add chinese/english documentation
* Make should shuffle = false in unittest
* Add python files to depends.

The reference return type causes ThreadParameterUpdater.cpp:123 seg fault under gcc5.4.

Change-Id: I7a1c155892722076a7cb48793b83d5ee525747d1

ISSUE=4602845

git-svn-id: https://svn.baidu.com/idl/trunk/paddle@1448 1ad973e4-5ce8-4261-8a94-b56d1f490c56

git-svn-id: https://svn.baidu.com/idl/trunk/paddle@1411 1ad973e4-5ce8-4261-8a94-b56d1f490c56

ISSUE=4586769

git-svn-id: https://svn.baidu.com/idl/trunk/paddle@1409 1ad973e4-5ce8-4261-8a94-b56d1f490c56

ISSUE=4586495

git-svn-id: https://svn.baidu.com/idl/trunk/paddle@1408 1ad973e4-5ce8-4261-8a94-b56d1f490c56