Error at TripletLoss forward()

[zosel260]

Hello. Thank you for sharing good works. I tried training on VisDrone2019-MOT-train dataset.

First, I built libs : DCNv2, correlation_package

Second, I generated dataset & path file using gen_dataset_visdrone.py

• dataset folder: /visdrone_fomt/train/images & labels_with_ids
• path file : data/visdrone.train

Third, Once I run train.py I faced an error at TripletLoss forward() of the loss.py

dist_an.append(dist[i][mask[i] == 0].min().unsqueeze(0))

and It is because sometimes targets = [all ZEROs] ( e.g, targets=[0 0 0]) Then, mask[i] == [ all TRUEs] and dist[i].[mask[i] ==0] is empty.

Could you give me any advice?

[michealangel]

may I ask you which repo and which configuration of DCNv2 did you use? I mean the torch version, cuda, Ubuntu. Thanks in advance