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

Training ProtoNet from scratch can only attain ~29.5% acc #2

Closed leodmel closed 5 years ago

leodmel commented 5 years ago

Hello, I can only attain 29.44% accuracy when running train_protonet.py by the default setting, however, ProtoNet should get about 49.42% accuary according to the paper, do you run train_protonet.py by its default setting or should I set some hyperparameters manually? Any idea would be appreciated.

Han-Jia commented 5 years ago

Hello, some default parameters are used for the pre-trained model. Please try the learning rate (lr) 0.001 or 0.002, and set step_size/gamma as 20/0.5.

If you run with the parameters above, you can get 1-shot 5-way performance at around 45%. The 49% performance shown in the ProtoNet paper uses the 30-way meta-training for the 5-way test.

leodmel commented 5 years ago

@Han-Jia Thanks a lot!