Closed kaixinbear closed 4 years ago
Hi and thank you!
No, this is not a normal phenomenon. Usually the axis and heading rise in accuracy pretty quickly, especially the axis which is intuitively easier.
Can I ask what modification did you make and does this occurrence happen without the change on your system?
Tnanks for reply. I use focal loss to replace ce loss, and I modify the loss.py from
if self.decomp_alpha:
loss_bbox_3d_init += float((loss_axis[torch.isfinite(loss_axis)].mean()
+ loss_head[torch.isfinite(loss_head)].mean()).item())*self.bbox_axis_head_lambda
if self.decomp_alpha and self.orientation_bins <= 0:
bbox_3d_loss += (loss_axis.mean() + loss_head.mean())*self.bbox_axis_head_lambda
to
if self.decomp_alpha :
bbox_3d_loss += float((loss_axis[torch.isfinite(loss_axis)].mean()
+ loss_head[torch.isfinite(loss_head)].mean()).item()) *self.bbox_axis_head_lambda
Would this modification affect the backprop of loss?
Thanks to your comment, I caught a minor bug in the config file that was uploaded in the initial commit. I have just updated it. It wrongly included the "orientation_bins" value to be set to a value > 0 and referenced a different model.
Can you check to see if this change solves your axis/heading acc?
I see your updates and the acc of head/axis converges to 1 normally. Thanks.
Hi, thanks for your great work . I train your project with mirror modification in network and loss. And I find something strange when seeing the training log. The axis and head accuracy is always around 0.50. It is the normal case that the acc of axis and head would converge to 1, right? I wonder to know is this two item around 0.5 a strange phenomenon? Thanks.