Closed VariantXYZ closed 1 year ago
Main fix is merged. Convolution is confused when all its geometric axes are gone, so we'll refrain from removing the last one (it will be re-expressed as a matmul anyway).
The nnef syntax one is on its way.
@kali , do you want me to make a separate issue for NNEF loading?
I think I fixed it, that was the second PR. Except it's not a load problem, you have to dump the network again, I have extended the syntax to support arbitrary strings as ids.
Ah, you're right. It works after updating.
Thanks!
I've been looking into https://github.com/breizhn/DTLN-aec, converted to ONNX using the very convenient tflite2onnx python package, and have found that it seems to spit out an error when trying to optimize (the model analyzes and runs fine otherwise).
You can convert it yourself, but I've attached the model in a zip file here.
2.zip
The node in question does not seem particularly egregious:
(512 512 512x1, it seems like the outer '1' is incorrectly being tacked on) but I may be missing something obvious.
Also, I wasn't sure if I had made a mistake, but when I tried to dump these as NNEF and optimize that way, I instead ran into a strange parsing error. Let me know if this is a separate bug and I'll go ahead and file it as well.