search

Sha-Lab / FEAT

The code repository for "Few-Shot Learning via Embedding Adaptation with Set-to-Set Functions"
MIT License
418 stars 84 forks source link

Hyperparameter training command for protonet with WRN backbone on miniImagenet dataset #54

Closed techNN closed 3 years ago

techNN commented 3 years ago

Hi, I appreciate your very insightful work as well as providing source code. Slightly related to Issue #24, I have been trying to recreate WRN pre-trained weights for miniImagenet that you had provided here for the protonet but can't seem to get the 61.4% performance you obtained in those results. Would you be willing to provide the command you used to train that protonet WRN model?

Thank you

Han-Jia commented 3 years ago

Hello,

For WRN, I use --batch_size 128 --lr 0.1 --dp_rate 0.3 --gamma 0.1 in pre-training. While in meta-training it uses the same command as with the ResNet backbone.

Given the pre-trained weights of WRN, please use the --fix_bn option to set the backbone to evaluation mode during training (which closes the dropout and stops updating the BN statistics) during the meta-training phase. We find --fix_bn is important for WRN.

techNN commented 3 years ago

Thank you, that is very helpful.