The backward compatibility of tract-onnx between 0.20.x and 0.21.x is broken

[cospectrum]

The model worked in version 0.20, but in version 0.21 the parsing failed.

scrfd.onnx.zip

Error: Failed analyse for node #204 "Resize_99" Resize

Caused by:
    0: Infering facts
    1: Applying rule inputs[2].shape[0] == {(inputs[0].rank) as dim}
    2: Impossible to unify Val(0) with Val(4).

[cospectrum]

The input is 1,3,640,640,f32

[kali]

Thanks for the report, I'll try to have a look

[kali]

I have merged a fix on main. Please be give it a shot. Also, be aware that tract Resize operator is weak and incomplete. It's a whole project of its own, with a lot of added complexity due to the multiple variants of its specification depending on operator sets. I just don't have the bandwidth.

[cospectrum]

model.run(inputs).unwrap() failed

[.../onnx/src/ops/resize.rs:11] ctx.onnx_operator_set_version = 11
[.../onnx/src/ops/resize.rs:11] ctx.onnx_operator_set_version = 11

called `Result::unwrap()` on an `Err` value: Evaluating #219 "Resize_99" Resize: output 0, expected 1,24,40,40,F32, got 1,24,20,20,F32 ...

[hturner08]

I also had to revert to 0.20 after running into a similar issue with wespeaker's diarization models. Inference with the unoptimized models fails:
called Result::unwrap()on anErr value: Input at index 0 has incorrect dtype or shape (got 1,82,80,F32 8.418448, 7.643861, 12.142844, 13.253984, 13.3040695, 12.930961, 12.339676, 11.143524, 12.737635, 13.239045, 13.041162, 13.224225..., expected to match fact B,T,80,F32)