Evaluator accumulate

[typhoonzero]

Fix #3808

[qingqing01]

def __create_var__ -> def _create_var.

[qingqing01]

看了下这个Evaluator里只有对多个mini-batch间的结果累加求平均。但对于PrecisionRecall这种多个mini-batch间结果统计并不是简单的累加求平均。

所以感觉如果在Python里抽象出Evaluator，用来统计多个mini-batch间的计算结果，需要抽象出一个Evaluator基类接口，然后有很多子类，AccuracyEvalutor, PrecisionRecallEvalutor等等。

[typhoonzero]

嗯，前面定义了avg_accumulate，是解决这个问题的。但的确没有考虑编译时。需要把avg accumulator等作为op，才可以使python定义的计算是编译时的。

[helinwang]

@qingqing01 @typhoonzero @wangkuiyi from my understanding the PaddlePaddle program is divided into compile time (currently Python) and runtime (C++): the Python code generates a protobuf which describes the computation, and the C++ code runs the computation description.

This implementation of the executor is in Python (uses scope which belongs to the runtime), I am not sure if it fits with the compile time and runtime concept. I think this requirement is very important for every OP implementation, @qingqing01 @wangkuiyi perhaps we need to make sure that everyone is on the same page.

[typhoonzero]

@helinwang you are right, I'll try to put this implementation to compile time.

[dzhwinter]

Currently, we can not create operator immediately out of Python Block and Program.

Paddle/python/paddle/v2/framework/framework.py

Lines 430 to 446 in 1f11f77

	class Program(object):
	def __init__(self):
	self.desc = core.ProgramDesc()
	self.blocks = [Block(self, 0)]
	self.current_block_idx = 0
	def __str__(self):
	protostr = self.desc.serialize_to_string()
	proto = framework_pb2.ProgramDesc.FromString(str(protostr))
	return proto.__str__()
	def clone(self):
	p = Program()
	p.desc = core.ProgramDesc(self.desc)
	p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())]
	p.sync_with_cpp()
	return p

May I just merge this work into codebase and let me polish it to fit in the refactorization code? : )
Our book porting work need this feature eagerly. @typhoonzero @qingqing01

[typhoonzero]

Sure, Thanks. Please also refer to some of my thoughts: #5157

[dzhwinter]

LGTM!
I just merge this work into codebase and let me polish it.
Our book chapters need this feature eagerly.
@typhoonzero @helinwang