ONNX Denoising models are invalid.

[FlorentGuinier]

Hi

We tryed running ONNX denoising model denoise_c3_ldr_float16.onnx and denoise_c9_ldr_f16.onnx from https://github.com/GPUOpen-LibrariesAndSDKs/RadeonImageFilter/tree/master/models using ONNXRuntime

It appears those models are invalids in term of ONNX specification:

The Resize nodes are indeed taking an fp16 tensor for the scale input.

However according to ONNX specification it should be fp32. https://github.com/onnx/onnx/blob/master/docs/Operators.md#Resize

Notes: I believe this is due to a bug in the package/tool that was used to produce those onnx file, exporting using updated package might solve.

Thanks Florent

[BenjaminCoquelle]

yes correct it was from the toolchain we used back then. Actually even in March this year we still got a fp16 tensor for scales. We will try again with very latest

Though I guess you tried the denoiser from this library itself as I know Unity uses it for their lightbaking

May I ask why you want to use it from onnxruntime instead?

[FlorentGuinier]

Hi, I'm working on Unity inference engine Barracuda and the goal here was to add the model to our tests where we typically compare our inference result to the one of ONNXRuntime for correctness test. This is in a context of an experiment by the lighting team there (Kasper in particular).

[FlorentGuinier]

I guess a good workaround for our use case would be to have the fp32 version of the model as we are looking for correctness here not performance.