Symbolic shape inference support

[jcwchen]

Current ONNX shape inference limitations

1. Will stop if it encounters symbolic dimensions. e.g., A (52) concat B (N2) = C (?*2)
2. Will stop if it bumps into some dynamic scenarios. e.g., Reshape to a dynamically computed shape

Work items

1. Extend shape inference with symbol generation and propagation
   • Extend graph-level shape inference
   • e.g., 5*2 + N*2 -> Create a new symbol M to represent its inferred shape
2. Partial data computation and propagation
   • Extend a new optional function PartialDataPropagationFunction just like existing TypeAndShapeInferenceFunction
   • Targeted operators: Add, Cast, Concat, Gather, Mul, Reshap, Shape, Slice, Size, Squeeze, UnSqueeze
3. (Independent) Propose a new ONNX Op to facilitate shape computation
   • There are examples that they need a few operators to extract certain dimension
   • To simplify these operators, add a new op Dimension for extracting a particular dimension value from the shape of input

I am doing 1. now and we plan to support it in ONNX 1.10.

cc @askhade @gramalingam

[askhade]

Closing this as all the identified work has been done. As more items come up in this area lets create a new issue