diyessi / onnx-builder Goto Github PK

Reorganize the directories according to https://packaging.python.org/en/latest/tutorials/packaging-projects/#packaging-python-projects

Remove the current kludge for optional outputs and have the builder determine which optional outputs are not being used by gathering names of inputs.

• Creating an exporter input should return a value that can be used as a placeholder.
• Exporter could have kwargs of input name and (type, shape)
• exporter.inputs[name] can be used as a value for that input

Variadic inputs (like Concat.inputs) and outputs (like Split.outputs) need to be accessible.

Need to also investigate the Sequence ops and data type.

Rather than dynamically creating wrappers, have a tool that generates updated python files for the wrappers so tools work.
Maybe grab the schema doc for the wrappers.

Can we use setattr and getattr to move more of the op classes into Node? For example, if we lists of the input names, attribute names, and output names (including info about which inputs/outputs are optional) Node could handle node_input_values, node_attributes and node_output_values making op definition a matter of providing strings, which we may be able to obtain directly from onnx.

There is no need for separate classes for each op; Node can have a member that contains node_parameter_descriptors, node_attributes and node_value_descriptors; possibly the onnx schema object for the op could be used directly since the descriptors and attributes are all derived from the schema.
Op constructors would be replaced with factory functions with the same name as the current classes that construct Node with the schema information object and the factory args -- inputs and attributes. These factory functions could be programmatically constructed from the schema.

Does opset downgrading work well enough to export as the current opset and then downgrade with ONNX tools rather than attempting to make ops know how to behave as different opsets?

Change Exporter to Graph and build the graph as wrappers are called.

Only create one Constant for a given constant value.

Separate Value and Node implementations so Value 0 is no different from any other Value.

• Use NumPy types
• Allow shape to be specified at build time like names can be