Skip to content

Update figure_iou.py#387

Open
salkhokhar wants to merge 1 commit into
meituan:mainfrom
salkhokhar:patch-1
Open

Update figure_iou.py#387
salkhokhar wants to merge 1 commit into
meituan:mainfrom
salkhokhar:patch-1

Conversation

@salkhokhar

@salkhokhar salkhokhar commented Aug 1, 2022

Copy link
Copy Markdown

As mentioned in issue (#319), sometimes during training the IOU loss gets stuck at nan. @qinhao14 suggested to use the function torch.nan_to_num() and it did solve the problem for me. And after multiple training sessions, no new problem seems to have arisen because of this.

As mentioned in issue [meituan#319] (meituan#319), sometimes during training the IOU loss gets stuck at nan. [@qinhao14 ] (https://github.com/qinhao14) suggested to use the function torch.nan_to_num() and it did solve the problem for me. And after multiple training sessions, no new problem seems to have arisen because of this.
@shensheng272

Copy link
Copy Markdown
Collaborator

Thanks for the pr.
Torch.nan_to_num() did make training more robustly on large batchsize and custom data. However most of our exps did not use this trick. All release model also did not use this trick.
So could you change code so torch.nan_to_num() check is off by default. Like store true training params

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants