Mmdetection: only one class, bug

Created on 24 Jan 2019  路  3Comments  路  Source: open-mmlab/mmdetection

when num_classes=1, g = torch.abs(input.sigmoid().detach() - target)
RuntimeError: Expected object of type torch.cuda.FloatTensor but found type torch.cuda.LongTensor for argument #3 'other'

Most helpful comment

num_classes needs to be set to 2 (such as person + bkg ) and reg_class_agnostic in bbox_head should be set to "True".
I successfully get the person detector by mmdet. That's a very good work!

All 3 comments

num_classes needs to be set to 2 (such as person + bkg ) and reg_class_agnostic in bbox_head should be set to "True".
I successfully get the person detector by mmdet. That's a very good work!

thank you very much

Excuse me!. Do you used retina_head to train only person???

Was this page helpful?
0 / 5 - 0 ratings

Related issues

FrankXinqi picture FrankXinqi  路  3Comments

qifei123 picture qifei123  路  3Comments

happog picture happog  路  3Comments

yangcong955 picture yangcong955  路  3Comments

fmassa picture fmassa  路  3Comments