search

stark-t / PAI

Pollination_Artificial_Intelligence
5 stars 1 forks source link

Should we also train from scratch besides using COCO trained weights? #49

Closed valentinitnelav closed 1 year ago

valentinitnelav commented 2 years ago

Showld we attempt some trainings from scratch besides using transfer learning for the YOLO architectures? I do not remember if Robin tried this already in his thesis. I am curious if for 300 epochs having the start advantage (transfer learning) of the COCO weights really makes a significant difference.

stark-t commented 1 year ago

@valentinitnelav According to some sources, using pretrained weight does not necessarily imporve results by a greate margin if by its own enough data is available. It can howerever imporve training time, e.g. less epochs needed and even offer better generalizability.

valentinitnelav commented 1 year ago

@stark-t that is helpful to know; if you happen to have some links at hand about this, could you put them here? This might be handy later.

stark-t commented 1 year ago

Using Pre-Training Can Improve Model Robustness and Uncertainty