linear is fine only if you have enough resolution, if most of your weights are close to zero, you risk quantizing all of them to 0 (or even worse to some non-zero value if you quantize linearly in min-max range)
of course, everything depends on the structure of the data to be quantized
but generally speaking, I'd pick a quantization scheme that lowers MSE any time
removing outliers seems dangerous, but I may be wrong
quantization-aware training seems like a great idea though
Neural network quantization
Moderators: hgm, Rebel, chrisw
-
- Posts: 2559
- Joined: Fri Nov 26, 2010 2:00 pm
- Location: Czech Republic
- Full name: Martin Sedlak
Re: Neural network quantization
Martin Sedlak
-
- Posts: 217
- Joined: Fri Apr 11, 2014 10:45 am
- Full name: Fabio Gobbato
Re: Neural network quantization
I have tried to use quantized weights in the training, only when I calculate the error of the network and seems to work. I have to try with int8 but with int32 works well.
Thank you!
Thank you!