Hi, everyone,
I have encountered the resize problem, that the output of snpe_resize operation varies a lot from that of Pytorch's interpolate. e.g. an upsample resize (13x13 -> 26x26).
Could you kindly help me with this issue? Is the upsample resize unsupported in snpe yet?
btw, I use the version of snpe-1.53.2, python-3.6.9, opset_version=10 (export onnx)
I'm not sure if the information I'm about to provide will be useful to you or not. But I've had issues with SNPE Resize bilinear as well. I think because when the layers is quantized and run on DSP (my assumption this is your setup), it cause the bilinear interpolation to be not what we expected. Instead, I switch to Nearest mode. In my experience, "nearest" doesn't degrade the model's performance, yet seemed to perform much better when quantized.
Thanks a lot for your reply. I try the "nearest" mode and it works!!! BTW, the bilinear interpolation output incorrect result ont only with qquantized model but also for the float mode. I found that the part of the interpolated pixels output incorrect but others right. This may be caused by their transformation matrix.