search

shule-deng / RFDNet

4 stars 1 forks source link

can you offer the model weigth for reproducing the result? #3

Open wushaoling540 opened 1 year ago

wushaoling540 commented 1 year ago

I can't reproduce the results provided in your paper, the results I produced by myself were much lower than the paper offered. For example, the results in the miniImageNet 5-way 1-shot setting were 64.41±0.85 in Acc and 69.28±0.81 in AUROC where were 66.40±0.82 and 71.91±0.78 respectively in the paper, and in the 5-way 5-shot setting were 80.78±0.59 in Acc and 76.21±0.72 in AUROC where were 81.19±0.60 and 82.29±0.56 respectively in the paper.

Could you please offer the model weight for reproducing the result?

wushaoling540 commented 1 year ago

Actually, I want to produce the result with ResNet 12.

shuledeng commented 1 year ago

I'm very sorry that my saved weights have been lost and I need to retrain, but I've been busy with my graduation thesis recently and it might take a while. In addition, my suggestion is that if you need to reproduce on ResNet12, you can directly train a ResNet12 version after adjusting the hyperparameters, which should have nothing to do with the results of ResNet18. Hope this helpful for you

wushaoling540 commented 1 year ago

Thanks for your reply.

shuledeng commented 1 year ago

You are welcome. Do you have any other questions? If not, I will close this issue