[WIP] panic should be a Scalar instead!

[NerdToMars]

Describe the bug

When converting an onnx model got panic as shown in the title.

Adds some debugging logs:

DEBUG burn_import::onnx::to_burn: Node outputs first : Some(Argument { name: "gather1_out1", ty: Tensor(TensorType { elem_type: Int64, dim: 0, shape: None }), value: None, passed: false })

ERROR burn_import::logger: PANIC => panicked at /root/.cargo/registry/src/index.crates.io-6f17d22bba15001f/burn-import-0.15.0/src/burn/ty.rs:200:9:
assertion left != right failed: name: gather1_out1; shape:None; kind:Int --???: Trying to create TensorType with dim = 0 - should be a Scalar instead!

The pannic is happening when try to convert the Onnx Argument type to burn Type.

Here is the onnx op info:

Some(Argument { name: "gather1_out1", ty: Tensor(TensorType { elem_type: Int64, dim: 0, shape: None }), value: None, passed: false })

To Reproduce

Expected behavior

Screenshots

Desktop (please complete the following information):

• OS: [e.g. iOS]
• Browser [e.g. chrome, safari]
• Version [e.g. 22]

Smartphone (please complete the following information):

• Device: [e.g. iPhone6]
• OS: [e.g. iOS8.1]
• Browser [e.g. stock browser, safari]
• Version [e.g. 22]

Additional context

[laggui]

Thanks for reporting!

Linking a related issue: #2699

Support for scalar tensors in ONNX import has been added progressively but the importer is not complete. The effort is mostly community-driven at this stage since we don't have enough bandwidth internally.

[NerdToMars]

Thanks for super fast response, I'd like to help, but probably need to take some time to look into this issue cuz I just start to use burn(thx for this nice crate it's awesome).

If possible would you mind sharing your thoughts about the direction of potential solutions?

[laggui]

If the issue really is with the output of your gather node that is a scalar, we need to expand the GatherNode implementation to support scalar outputs.

The codegen in forward(...) only works for a tensor output.

That also means that the conversion should accept scalar or tensor types (which is probably the current point of failure given the panic message).