MendelXu/SAN

There is also a question about "emsemble"

sunwhw opened this issue · 2 comments

sunwhw commented

There is also a question about the emsemble, what should the weight be set to emsemble the origin san model and the model fine-tuned on coco stuff? or could you tell me the range the weights should be set?

w = float(os.environ["EWEIGHT"])
sem_seg = torch.pow(sem_segs[0][:min_cls],w)*torch.pow(sem_segs[1][:min_cls],1-w)
sunwhw commented

and when you fine-tune on the coco stuff, is your model totally samed as the SAN?

MendelXu commented
  1. In most situation, the weight is set to 0.5.
  2. Yes, the model is exactly the same network as SAN.
  • 👍1