search

meituan / YOLOv6

YOLOv6: a single-stage object detection framework dedicated to industrial applications.
GNU General Public License v3.0
5.72k stars 1.03k forks source link

QAT with different teacher model #461

Closed haritsahm closed 2 years ago

haritsahm commented 2 years ago

From the paper and code implementation, QAT + distillation will use its own configuration with different precisions (fp32 for teacher and int8 for student). What if the teacher use different architecture, i.e YOLOv6L (teacher) -> YOLOv6-S/T (student), is there any benefits from this approach?

lippman1125 commented 2 years ago

@haritsahm I tried to use M/L as a teacher, but didn't get mAP improved, I guess too much gap exists between large-sized teacher and small-sized student.

haritsahm commented 2 years ago

@lippman1125 I see, thank you for the explanation