Enrico Civitelli

Hello all, I'm trying to run YOLOv3 using float16 instead of float32 at inference time, I have changed both network and data type as described here. Unfortunately, the script throws...

Hi @zhreshold, It gives the same exception also changing that value to "float16". It seems that somewhere this implementation of YOLO casts something to float32 by default. Have you ever...

Hello all, I have probably found a way to run YOLOv3 in float16. Basically, the problems are some hidden hardcoded float32 tensors.. I will put my "corrections" here asap.

Yes, I have successfully trained YOLOv3 with a reduced precision (float 16) some time ago. The problem is in the forward function: if I'm not wrong there are some operations...

Hello Cao_enjun, Excuse me, I can't understand the problem. Can you please provide a more detailed explanation? Is the same problem present in Adversarial Training (AT)? Thank you, @enry12