Optimized Clipping
Current clipping method of the PTQ algorithm use the maximum values of the unit outputs. A better way to procede would be to reduce this range in order to increase the accuracy of the most frequent values. To do so we want to compute the histograms of the activation values and etablish the best threshold according to a metric. In a first time, we will use MSE (Mean Square Error), and in a later one KLD (Kullback-Leibler Divergence).
EDIT: Solved by !9 (merged)
Edited by Benjamin Halimi