[MXNet Forum] [Gluon] Help with SSD SmoothL1 metric reporting NaN during training

[Deniz Kenan Kilic via MXNet Forum <mx...@discoursemail.com.INVALID>]

Dear Lee, I have the same problem for trying fine tuned ssd code with my custom dataset. While I was preparing .lst file I wrote the lines as following for images that have no objects:
idx 4 5 512 512 -1.0 -0.001953125	-0.001953125	-0.001953125	-0.001953125	img_path

I added class id as -1 and added bbox as [-1, -1, -1, -1] (above are normalized). How did you solve the issues with ground truth label, especially for images that have no objects? By the way I tried also the following format for no objects:
0	4	0	512	512	img_path
However it gives error about labeling in ssd code. Thanks so much for any help!





---
[Visit Topic](https://discuss.mxnet.io/t/help-with-ssd-smoothl1-metric-reporting-nan-during-training/4666/6) or reply to this email to respond.

You are receiving this because you enabled mailing list mode.

To unsubscribe from these emails, [click here](https://discuss.mxnet.io/email/unsubscribe/3ec501e4bc75827aa27a18e8f477180e3dd0cc9368a309294bbbd8433fffa545).