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

Pretraining hyperparameters for CUB dataset #76

Closed mayug closed 2 years ago

mayug commented 2 years ago

Hi,

Could you please provide the hyperparameters used for pertaining convnet and res12 on the CUB dataset?

Also, what were the best val_dist acc and val_sim acc you got for convnet and res12 on the CUB dataset in the pertaining stage?

Thanks and Regards Mayug

Han-Jia commented 2 years ago

Hi,

For ConvNet, please try python pretrain.py --lr 0.001 --batch_size 256 --max_epoch 500 --backbone_class ConvNet --schedule 300 350 400 450 --gamma 0.1 --dataset CUB --query 15 [Dist] best epoch 141, current best val acc=0.1758 (over all meta-val classes) [Sim] best epoch 121, current best val acc=0.1942

For ResNet, please use lr = 0.1, epoch = 600, schedule = 400, 500, 550, 580, gamma=0.1 we only record the 1-shot 5-way results based on learned embedding with the cosine classifier, it achieves almost 73%.