Created by: liym27

used for CI check https://github.com/PaddlePaddle/Paddle/pull/21522

1. add interface to get default value of attributes

m.def("get_op_attrs_default_value", [](py::bytes byte_name) 
           -> paddle::framework::AttributeMap{...}

2. add file check_op_desc.py to print op desc information and compare op desc

https://github.com/PaddlePaddle/Paddle/pull/21613 move get_all_ops_desc from check_op_desc.py to print_op_desc.py

(1) Print all ops desc in dict

run python check_op_desc.py > OP_DESC_DEV.spec to get ops desc of branch develop run python check_op_desc.py > OP_DESC_PR.spec to get ops desc of branch PR

after #21613 : run python print_op_desc.py > OP_DESC_DEV.spec to get ops desc of branch develop~~ run python print_op_desc.py > OP_DESC_PR.spec to get ops desc of branch PR~~

(2) Compare op_desc files generated by branch develop and branch PR and print error message.

run python check_op_desc.py OP_DESC_DEV.spec OP_DESC_PR.spec > error_message

after #21613 : run python print_op_desc.py OP_DESC_DEV.spec OP_DESC_PR.spec

3. The following are non conformances

If the modification of op from branch PR conforms to the specification, the error_message is empty. The modification of Ops will cause that the new version inference fails to load model trained by the old version.

• for ops changed in branch PR:

  • add Input/Output but it's not dispensable
  • delele Input/Output
  • modify Input/Output including duplicable, intermediate or dispensable
  • add Attr but its default value is not set
  • delele Attr
  • modify Attr including the type, default value or generated

• for ops deleted in branch PR:

  • No comparison

• for new ops only added in branch PR:

  • No comparison