Is the supervision of losses_pts_refine necessary?

microsoft / RepPoints

Represent Visual Objects by Point Sets

MIT License

589 stars 75 forks source link

Is the supervision of losses_pts_refine necessary? #34

Closed zcy5417 closed 4 years ago

zcy5417 commented 4 years ago

I remove the losses_pts_refine in the loss_dict_all, and want to supervise the bbox regression only by the losses_pts_init, but find that in the training process, loss go to NaN after several hundread iters. Is the loss_pts_refine necessary?

zcy5417 commented 4 years ago

sorry i made a mistake