I have tried running our organization's countdown using snpe-net-run on various gadgets, in computer CPU mode.
I tried four gadgets. The first is PC, the other is snapdragon 636, the third is snapdragon 865, the fourth is snapdragon 888. I have a similar result on PC, snapdragon 636 and snapdragon 888. Anyway, the result is on snapdragon 865 is completely different. Also, I additionally look at the result of each layer by adding -- examine. I noticed that the consequence of the main layer is not correct.
Our first layer is:
self.first_conv = nn.Sequential(
nn.Conv2d (3; 24; 3; 2; 1; gap = False),
nn.BatchNorm2d(24),
nn.ReLU (in place = True),
)
The resulting tensor has 98,304 absolute qualities and the 88,512 initial qualities are equivalent in all devices. Also, the excessive quality on the snapdragon 865 is not the same as the results on different gadgets.
Can anyone let me know how to take care of this problem? Cordially. Since our application needs to be sent to the snapdragon 865, we should get the correct results for this gadget.
Dear customer,
You mean the execution results exist difference while you perform on variant devices?
How do you verify the model accuracy from SNPE SDK within same commands?
BTW, what's SNPE version you used?
BR.
Wei