Remaining tasks for fluid

[reyoung]

We complete the most jobs for fluid release, but still there are several tasks remains.
tasks.dot.txt

There are two colors in each block, aquamarine and white. The tasks in the white blocks is under developing right now. We need help to complete the tasks in the aquamarine blocks.

NOTE: The tasks and graph of dependencies could be flux.

There are five major parts of tasks:

1. Polish our Python API. Make the high level Python API as same as our V2 API.

• Transform training ProgramDesc to testing need to handle dropout/batch_norm operator in testing mode #5733 support testing when training and handle dropout and batch_norm operator in testing mode #5734
• test while training Support testing while training #5814 support testing when training and handle dropout and batch_norm operator in testing mode #5734
• Basic Performance tunning
  • Performance tunning guide documentation Feature/cpu profiling #5895
• Training data IO API desgin
  • Executor can use numpy tensor as its input/output "add asnumpy interface" #5620
  • Executor can use numpy list for LoD i/o
• Parameter Attribute Design and initializers Parameter Attribute Design and initializers #5819
• Evaluator design
  • Simpilify and enhance evaluator.py #5825
• Enhancement for fluid API
  • Restore param info when Program.clone() Restore the information of Parameters in Program.clone() #5871
  • Unify the fluid imports into paddle.v2.fluid Unify the fluid imports into paddle.v2.fluid #5876
  • A Better way to switch default program Need a better way to switch default program #5929
• Paddle V2 Proxy Layer In next release.
• Make API function parameter name consistantly
• Polish C++ operators
  • Operator should support many data types, not only float. Sum operator should support int, int64 #5827
• Machine Translation Demo
  • neural machine translation training complete neural machine translation trainer #6400
  • predicting

2. Complete our last two chapter in our book.

• 07/Label semantic roles 07/Label semantic roles #5798

3. Synchronize our training accuracy with previous Paddle and other frameworks.
   • We train same data with same configuration and same initialization. We should make fluid produce the exactly same result with previouse Paddle or other frameworks. Exactly means the epsilon should be less than 1e-4.
4. Polish Op & Layer documentation.
   • The most crucial task is generate Op & Layers documentation HTML files, and publish it in our staging sever.
5. Releasing new docker images & pip packages
   • Also run the regression tests.

[reyoung]

Done.