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

reason behind good performance of Convnet over ResNet for noisy image #64

Closed kumam92 closed 3 years ago

kumam92 commented 3 years ago

I am using your FEAT model for my own dataset which is quite noisy. I found that when I use FEAT model ResNet 12 is better than ConvNet but when I use same prototypical network as in your code I found that ConvNet outperforms ResNet12 by 3%. What could be the reason for that FEAT best accuracy is through deeper network and prototypical network works better with shallow network?

Han-Jia commented 3 years ago

Hi,

One reason could be that the Transformer compensates for the noise in different tasks. So when using the ProtoNet, the deeper network could memorize the noise in the base class set and cannot generalize to the UNSEEN data.