Transforming scaling factor in shift and multiplier

Problem description

I don't know wich function should I use to compute the shift and the multiplier for the scaling. In the Simple case of a fully_connected with an input of size 10 and an output of size 5, all of the result are overload

  ------------------ OUTPUT ----------------------
 Input offset : 0
 Output offset : 0
 Input dims : n = 1, w = 1, h = 1, c = 10
 Filter dims : n = 10, c = 5
 Output dims : n = 1, c = 5
 Bias dims : c = 5
 Quantization params : multiplier = 131072, shift = 17
 Buffer size : 0
 Input Data :
[ -48 , -118 , -128 , -73 , -45 , -94 , -1 , -50 , -85 , -100 ,]
 Output Data :
[ 127 , -128 , -128 , 127 , 127 ,]

Reproducible

Reproductible in this commit : b183a2d6

added StatusWIP label

assigned to @bhalimi

assigned to @wboussella and unassigned @bhalimi

@wboussella Please fill the content of this issue with your ongoing work for adding quantized network to the export.

changed the description

changed title from PTQ to Transforming scaling factor in shift and multiplier

changed milestone to %aidge - May 2024

closed

removed StatusWork in Progress label