search

lenscloth / RKD

Official pytorch Implementation of Relational Knowledge Distillation, CVPR 2019
390 stars 49 forks source link

Can you release your trained model on cub200 ? #8

Open UnrealPawn opened 5 years ago

UnrealPawn commented 5 years ago

Hi @lenscloth ~ I can't get the same recall score on dataset cub200 emm....... My test: Teacher : resnet50 embedding : 512 batchsize: 128 sample: distanceweight and the result is 61.0 (very close to your score 61.24) But...student resnet18 is low...(58.74)without L2normal , dist_ratio=1 and angle_ratio=2

So can you release your trained model on cub200 ?(teacher and student model ) Any help would be deeply appreciated!

lenscloth commented 5 years ago

@UnrealPawn What was the embedding size of the student? If the embedding size was set to 64, the score (58.74) is same as the value in the paper (58.68).

If you wish to boost the student's performance more than teacher. I would recommend you to use large embedding size (e.g, 512) or a large backbone (e.g, resnet50).

Still, I will upload the trained models in the repo soon.