Skip to content

Fix the NaN bug

Benjamin Halimi requested to merge bhalimi/aidge_quantization:DevPTQ into main

Previous version of the PTQ implementation could result in a model producing NaN values. This was due to a poor memory initialization obtained by doing a multiplication by zero. Thus, rescaleTensor(tensor, 0) is now replaced by fillTensor(tensor, 0) which solves the problem.

Merge request reports

Loading