Closed leehangyu closed 4 years ago
Can you please bit a bit more precise about what the problem is? Where was the problem previously? In ONNX? The ONNX to NNEF converter does support bilinear upsampling. By the way, as I see from the ONNX documentation, the upsample operator has been deprecated in opset version 10. What changed in version 11?
Sorry for not being specific.
Let me show you how they are different.
This shows the result of upsampling bilinear in opset_version_9
This shows the result of upsampling bilinear in opset_version_10
This shows the result of upsampling bilinear in opset_version_11
as you can see, only opset_version_11 returns correct result.
Because of this, my model converted with opset_version_9 causes a critical difference between original and converted one.
To me, this seems to be an incompatibility issue between PyTorch and ONNX (runtime). What is the problem related to NNEF? Is there any error when you try to convert it to NNEF?
The only version I can convert it with is opset version 11 as you can see from previous screen shots, but NNEF does not support 11.
That's why I am asking for when NNEF supports opset version 11, there was warning show that NNEF does not support opset version 11? So I am wondering If there is any problem doing it with NNEF
Oh, by the way I figured out I do not need to interrupt the intermediate stage unless layers have weight, since upsampling layer has no weight I do not need to change the algorithm but pass the parameters.
I am sorry I didn't know it, I thought I had to correct exact match even intermediate stages but It is not.
The only upsampling layer I need to change is upsampling bilinear in TF-LITE model.
Thank you for reply and help.
Hi
There was a problem converting a upsample bilinear layer before opset_version 11 in onnx,
But it had been fixed it by version 11,
I was wondering if you are planning to support onnx opset version 11
Thanks