Open yahuuu opened 2 years ago
You can do this by modifying the mode of Conv Op in def convert_conv of this tool.
I have not tested the performance of this modification, I am looking forward to hearing your test result whether it does improve the speed.
I converted onnx to h5, but found that the padding property in the convolution is valid, how to modify the property to same. Whether this property affects the speed when post quantized model runs on the SnapDragon 855 DSP.