🚀 Feature

• Simplify the symbolic_sizes_ property of TensorType API. TensorType is a class to represent a tensor type in TorchScript IR. TensorType tracks various properties of a tensor such as tensor shape, strides, device, requires_grad, etc. TorchScript IR consists mostly of Node and Value . Nodes denote operations that produce outputs (Values) . Each value is statically typed with Type and that type can be TensorType .
• Currently, an unknown size is expressed with VaryingShape that wraps around a vector of c10::optional<int64_t>s
• We recently added ShapeSymbol class to represent size and values and it has its own way of expressing unknown values`
  • ShapeSymbol now provides a special symbol (-1) which we can use to denote unknown symbols (https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/core/jit_type.h#L389)
• In short, we would like to switch from using optionals to denote unknown symbols to using that special symbol.

Motivation

• The outcomes we hope for
  • The reduction in the amount of boxing/unboxing
  • Avoiding having more than one way to express the same thing.

Pitch

• We would like to add unknownSymbol to denote unknown/varying shape
• We would like to use unknownSymbol in lieu of c10::optional<ShapeSymbol>
• We would like to replace VaryingShape<ShapeSymbol> with just c10::optional<std::vector<ShapeSymbol>>

Alternatives

Do nothing n and suffer from the current design.

Additional context

For questions, ping @Krovatkin

cc @suo