ONNX conversion: Only tensor input is valid Argument

[AdrianEddy]

Describe the bug
When trying to convert onnx model to Burn, I'm encountering this panic:

  WARN burn_import::onnx::dim_inference: Must implement dimension inference for Slice
  WARN burn_import::onnx::dim_inference: Must implement dimension inference for Slice
  ERROR burn_import::logger: PANIC => panicked at burn-import-0.12.1\src\onnx\dim_inference.rs:311:9:
  Only tensor input is valid Argument

  --- stderr
  thread 'main' panicked at burn-import-0.12.1\src\onnx\dim_inference.rs:311:9:
  Only tensor input is valid Argument

The node_input on that line is { name: "cast2_out1", ty: Scalar(Int64), value: None, passed: true }

The model I'm trying to convert is gmflow-scale2-regrefine6-mixdata-train320x576-4e7b215d.onnx

I also have alternative onnx model exported by someone else, which errors with a different message, might also be worth looking into: gmflow-scale1-mixdata-train320x576-4c3a6e9a_1x3x480x640_sim.onnx

To Reproduce

1. Create a build.rs with:

fn main() {
    ModelGen::new()
        .input("gmflow-scale2-regrefine6-mixdata-train320x576-4e7b215d.onnx")
        .out_dir("model/")
        .run_from_script();
}

2. Watch it fail

Expected behavior
I expected it to finish conversion successfully

Desktop (please complete the following information):

• OS: Windows 11, Rust 1.77.1, Burn 0.12

Additional context
Model exported using torch.onnx.export from this repo

[laggui]

According to the line that panics it seems to be due to an Unsqueeze op.

Looks like this op is not full supported yet (#600).

Do you know what else is missing for the full support @antimora ?

[antimora]

Thanks, @AdrianEddy, for filing this issue. This helps us prioritize the ONNX ops.

OK. I have checked the ONNX file and it contains lots of nodes (> 10K) but with a few types:

1. Add
2. BatchNormalization
3. Cast
4. Concat
5. Constant
6. ConstantOfShape (CURRENTLY NOT SUPPORTED)
7. Conv
8. Div
9. Expand (CURRENTLY NOT SUPPORTED)
10. Gather
11. Identity
12. Mul
13. Relu
14. Reshape
15. Shape
16. Slice (CURRENTLY NOT SUPPORTED)
17. Sub
18. Tile (CURRENTLY NOT SUPPORTED)
19. Unsqueeze

Unsupported in ONNX import currently:

1. ConstantOfShape
2. Expand
3. Slice
4. Tile

The full list of supported ops: https://github.com/tracel-ai/burn/blob/main/crates/burn-import/SUPPORTED-ONNX-OPS.md

These missing ops came up recently in this issue as well: #1544

The reason why Unsqueeze fails is because it is referencing ConstantOfShape which isn't supported at the moment.

Tagging @nathanielsimard so he's aware of missing ONNX ops.

[antimora]

If we handle #1544 fully then this ticket will get resolved too.

[antimora]

I fixed unsqueeze scalar (#1690) issue but it's blocked on converting Gemm node to linear module:

[burn-import]% cr ./clip_image_model_vitb32_batchsize1.onnx ./out

....

DEBUG burn_import::onnx::from_onnx: output name: 
/visual/transformer/resblocks/resblocks.0/attn/Reshape_3_output_0
DEBUG burn_import::onnx::proto_conversion: Converting ONNX node with type "Gemm"
DEBUG burn_import::onnx::from_onnx: renaming node "/visual/transformer/resblocks/resblocks.0/attn/Gemm"
ERROR burn_import::logger: PANIC => panicked at crates/burn-import/src/onnx/dim_inference.rs:122:46:
called `Option::unwrap()` on a `None` value
thread 'main' panicked at crates/burn-import/src/onnx/dim_inference.rs:122:46:
called `Option::unwrap()` on a `None` value
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
[burn-import]%