failed to get the same performance

damo-cv / Ada-NETS

This is an official implementation for "Ada-NETS: Face Clustering via Adaptive Neighbour Discovery in the Structure Space" accepted at ICLR 2022.

MIT License

33 stars 3 forks source link

failed to get the same performance #13

Closed zzzzzero closed 2 years ago

zzzzzero commented 2 years ago

Thank you for your great work. I run the train and test scripts on the ms1m dataset these days,but I got poor performance. Here are my result on ms1m part1_test: Bcubed fscore:0.6264, parwise fscore 0.4994. I wonder which step went wrong，could you share the training log? When I run the train_GCN script ,the last loss is: loss:1.9265 bclloss_pos:0.007 bclloss_neg:1.919

Thomas-wyh commented 2 years ago

Don't Worry. Can you tell me trainmre and testmre result of adaptive filter after training?

zzzzzero commented 2 years ago

trainkdiff 5.21 testkdiff 8.04 trainmre:0.11 testmre:0.32

zzzzzero commented 2 years ago

After I change the version of dgl ,the performance is normal.