Only tensor input is valid when converting ONNX model

[BinaryWarlock]

Describe the bug

ERROR burn_import::logger: PANIC => panicked at .../burn-import-0.12.1/src/onnx/dim_inference.rs:403:9:
  Only tensor input is valid

Which is in conv2d_update_outputs.

Adding extra printing, it seems to be taking in a Scalar(Float32). I don't know where that's from.

To Reproduce
Using onnx-clip model which you can grab at https://lakera-clip.s3.eu-west-1.amazonaws.com/clip_image_model_vitb32.onnx (about 336 MiB)

1. Fix the input with python -m onnxruntime.tools.make_dynamic_shape_fixed --dim_param image_batch_size --dim_value 1 clip_image_model_vitb32.onnx clip_image_model_vitb32_batchsize1.onnx
2. Try importing with burn-input build script
3. Receive error

[antimora]

I will investigate next week.

[antimora]

The immediate issue is related to casting op, which will be fixed by this PR #1634

Once we merge outstanding PRs related to ONNX, I'll recheck it again.

I also noticed that our Transpose operator is not correctly implemented. Filed a ticket: #1642

[antimora]

OK. We made progress but I found another bug that's blocking.

[BinaryWarlock]

@antimora Do you happen to know which ops this model is blocked on currently?

[antimora]

@antimora Do you happen to know which ops this model is blocked on currently?

I fixed a bunch of limitations we had (mostly related to an older OpSet version) and I added a few Ops but we still lack the following.

We are missing these implementations (mentioned in #1714)

1. ConstantOfShape
2. Gather

We need to fix this

1. Reshape (needs to accept input for the shape).

The whole list of Ops used by the ONNX file.

1. Add
2. Cast
3. Concat
4. Constant
5. ConstantOfShape
6. Conv
7. Div
8. Gather
9. Gemm
10. MatMul
11. Mul
12. Pow
13. ReduceMean
14. Reshape
15. Shape
16. Sigmoid
17. Softmax
18. Split
19. Sqrt
20. Sub
21. Transpose
22. Unsqueeze

Please note, we might still discover some issues once we implement the missing ops and fix Reshape.