I have some ONNX models that convert to DLC and quantize really well (meaning they run fast and keep high accuracy) but I have some that don't. Do you have any whitepapers or anything on what layer types work well, and what to avoid so that the DLC and quantization work well? Also does the DLC conversion try and leverage the 4 threads on the DSP or do we have to design the model in a way that maximises parallelism potential? Any whitepapers on that topic?
Losing accuracy on DLC conversion
Posted: Mon, 2022-04-11 14:13
I think we've narrowed down the problem to the use of blinear interpolation. Without that things work better.