Python Generate OpCreation Methods by OpProto #2893
Conversation
All OpCreation method are generated by `create_op_creation_methods::__bootstrap__` method, and stores in `op_creations` object and its methods. There are three parts to implement this feature. 1. Get all registered `OpProto` from C++ side. It is implemented in `get_all_op_protos` method. 1. Create a function to convert `kwargs` to `OpDesc` base on each op's `OpProto`. The `OpDescCreationMethod` class. 1. Convert `OpProto` to `docstring` by `get_docstring_from_op_proto` method. All three methods are unit tested. The `__bootstrap__` just combines them together and create a method in runtime. For details, please reference the doc string in `create_op_creation_methods.py` and the unit test `test_op_creation_methods.py`.
reyoung
force-pushed
the
feature/op_creation_methods
branch
from
13ec575
to
9e0c680
Compare
paddle/framework/op_registry.h
Outdated
| @@ -222,6 +235,17 @@ class OpRegistry { | |||
| }; | |||
|
|
|||
| private: | |||
| static void AssignTempVariable(OperatorBase* op) { | |||
There was a problem hiding this comment.
Assign is not a good name, maybe GenNameForTempVar?
| for (auto& attr : op_desc.attrs()) { | ||
| op->attrs_[attr.name()] = AttrTypeHelper::GetAttrValue(attr); | ||
| } | ||
| op_checkers().at(op_type).Check(op->attrs_); | ||
|
|
||
| //! Convert Temporary variable name to an unique variable name. | ||
| GenerateTempVariableName(op.get()); |
There was a problem hiding this comment.
因为temp var name是c++这边自动生成的，所以python那边就管不了这些Var了是么，这些Var应该如何创建，被谁创建？
There was a problem hiding this comment.
Python这边会在运行Op之前扫描这个Op的输入和输出，进而创建对应的variable
| @@ -39,6 +39,13 @@ using OperatorPtr = std::shared_ptr<OperatorBase>; | |||
| */ | |||
| class OperatorBase { | |||
| public: | |||
| /// If a variable is a empty variable, that name will be used. | |||
There was a problem hiding this comment.
should add some more comment about what is empty variable and temp variable
| protostrs = core.get_all_op_protos() | ||
| ret_values = [] | ||
| for pbstr in protostrs: | ||
| op_proto = op_proto_pb2.OpProto.FromString(str(pbstr)) | ||
| ret_values.append(op_proto) | ||
| return ret_values | ||
|
|
||
|
|
||
| class OpDescCreationMethod(object): |
There was a problem hiding this comment.
class OpDescCreator(object):
def create(self, *args, **kwargs):
...
return op_desc I think above is better for reading and understanding code.
There was a problem hiding this comment.
A Function object(https://en.wikipedia.org/wiki/Function_object) or functor is a very common way for many programming languages.
Using Function object can hold internal states and make function implementation into small pieces and more readable.
| @@ -17,6 +16,7 @@ | |||
|
|
|||
| class Optimizer(object): | |||
| def __init__(self, **kwargs): | |||
| import py_paddle.swig_paddle as swig_api | |||
There was a problem hiding this comment.
what is the purpose of this change? to speed up the importing?
There was a problem hiding this comment.
The same reason in #2864
|
|
||
| class TestOpCreations(unittest.TestCase): | ||
| def test_all(self): | ||
| add_op = creation.op_creations.add_two(X="a", Y="b", Out="z") |
There was a problem hiding this comment.
creation.op_creations 感觉应该做一个alias
比如 ops 或者 operators，不过这个可以后边再搞
All OpCreation method are generated by
create_op_creation_methods::__bootstrap__method, and stores inop_creationsobject and its methods.There are three parts to implement this feature.
OpProtofrom C++ side. It is implemented inget_all_op_protosmethod.kwargstoOpDescbase on each op'sOpProto. TheOpDescCreationMethodclass.OpPrototodocstringbyget_docstring_from_op_protomethod.
All three methods are unit tested. The
__bootstrap__just combinesthem together and create a method in runtime.
For details, please reference the doc string in
create_op_creation_methods.pyand the unit testtest_op_creation_methods.py.