Some issues in the `framework.py`
Created by: JiayiFeng
Last weekend I tried to implement the trainer.test() in our newest design (#10248 (closed) ). During this, I noticed that a few functions and interfaces of framework.py are implemented in quite twisty ways. Some of them only make users confused while the others may result in serious bugs. Since framework.py is the pivot of our Python API, I propose to polish framework.py first before further development on Python API.
Here are some issues I have noticed in the framework.py:
-
VariableandParameterare using**kwargas it's__init__inputs. With**kwarg, users are likely to be confused about what is exactly required to create aVariableandParameter.
- Some code in the
__init__ofVariableis used to handle the situation that the variable to be created has existed. In this part of code, we check each attribute ofVariableone by one to make sure the existing variable is consistent with the one to be created, which makes the code extremely long. Is there any way to simplify this part?
- The function of
var()is opposite inBlockandScope. If the queried variable not exists, thevar()ofBlockwill throw an exception while the one ofScopewill create it.
The var() of Block:
https://github.com/PaddlePaddle/Paddle/blob/f43b71b242467d665c134262c2b7167cef622757/python/paddle/fluid/framework.py#L717-L723
The var() of Scope:
https://github.com/PaddlePaddle/Paddle/blob/f43b71b242467d665c134262c2b7167cef622757/paddle/fluid/framework/scope.cc#L51-L59
- The
clone_variable()ofBlockdoes not clone all attributes of variables correctly.capacity,error_clipandstop_gradientare omitted. And why all clonedVariable'spersistableare True?
- In the
Program'scloneinterface. The newProgramis created with the old one's C++ desc instead of the Python object itself. However, some information only exists in Python(e.g., error_clip, stop_gradient, ...). It mean aProgramcannot be cloned entirely with this interface.