Question: Unsigned Quantization

[mahdi259]

Hi
Is there any way to specify unsigned quantization in QuantIdentity or QuantLinear?
Thanks

[capnramses]

Hi @mahdi259
I think for QuantIdentity the answer is just to use Uint8ActPerTensorFloat when initialising QuantIdentity with optional argument act_quant=Uint8ActPerTensorFloat instead of the default int8 - which you can see here

brevitas/src/brevitas/nn/quant_activation.py

Line 106 in 2004568

class QuantIdentity(QuantNLAL):

For weights in e.g. QuantLinear I think it would be less common to use unsigned integers, so I don't think there's a datatype available built-in, but you could experiment with that by extending the Int8 weight data type

brevitas/src/brevitas/quant/fixed_point.py

Line 24 in 2004568

class Int8WeightPerTensorFixedPoint(NarrowIntQuant,

I hope that helps!

[mahdi259]

Thanks @capnramses . That's it. One more question. I have read ULPPACK paper that performs inner product of unsigned numbers, i.e. unsigned inputs and weights. I agree that it's odd to have unsigned weights. Do you know how they use ULPPACK in networks like ResNet?
Thanks

[capnramses]

I haven't read that paper I'm sorry! I think that might be off topic for the repo, so I'm going to close this Issue.