01-23-2020 10:26 PM
I have tensorflow model of Densenet. When I am running this on ZCU102, I get the following error:
[DNNDK] Invalid pixel value of input tensor: 19840
[DNNDK] Please check if decent tool produces correct quantization info.
I have checked the input image and the maximum pixel value in it was 221.
How could the DNNDK throw such error that the pixel value is 19840.
When I tried normalization by dividing with 255, it still says "Invalid pixel value: 128". When I print the Mat image all I could see are 1's and 0's.
02-25-2020 07:22 PM
Hi @jbond1007 ,
The input image pixel data need to be pre-processed before sending to DPU, you need to make sure that the pre-processing is just the same as when doing training, and also the input_scale value need to be multiplied to match the quantized network.
like code example here:
In this example the data is normalized in training, so it is done in ARM code.
Hope this can help.
06-29-2021 06:57 AM
My input_scale read from dnndk is 128, if I multiply 128 it will say: Invalid pixel value of input tensor: 5782.
But if I don't multiply the scale, it will say: Invalid pixel value of input tensor: 128.
if I change the scale to 0.9999, it will become normal.
So I want to ask that if dnndk (vitis ai) change the weight to +128 instead of +127 ?
This is my pre process: