Question about onnx exporting a model which uses torch.rfft() in the forward function. How would this work as ONNX opset doesn’t support FFT yet.
Do let me know. Thanks!
Question about onnx exporting a model which uses torch.rfft() in the forward function. How would this work as ONNX opset doesn’t support FFT yet.
Do let me know. Thanks!
I think you’ll just get a funny-looking operator (aten::…) in ONNX.
The other question is, of course, which backend would than execute your operator.
Best regards
Thomas
Would use x86_64 CPU to run this using onnxruntime.
From a cursory look, it seems to be a contributed operator shown as com.microsoft.rfft
there. Maybe you can pretend that aten::rfft is a custom op and register a custom ONNX operator for it.
Best regards
Thomas