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

Intitialisation vs final weights #14

Closed willwilliams closed 4 years ago

willwilliams commented 5 years ago

I think your provided weights for the initialisation are mixed with your final weights.

The README states: 1) Pre-trained weights are here - https://drive.google.com/open?id=14Jn1t9JxH-CxjfWy4JmVpCxkC9cDqqfE 2) The learned models are here - https://drive.google.com/open?id=1ZjkiEJh_96VYNWCOXUGsPuesLaFzV_z9

For MiniIMageNet-Res-5-Shot-5-Way.pth at the very least these appear to be the wrong way round.

Can you confirm?

willwilliams commented 5 years ago

Also, it appears the weights provided don't include those for the transformer. Is this expected?

Han-Jia commented 5 years ago

Hi,

To evaluate the model with the final weights, please also set the parameters like "shot" and "model" in the script. For the MiniImageNet with 5-shot 5-way, it is expected to get accuracy around 78%.

For the initialization weight, it only contains the parameters of the backbone. So the pre-trained weight only relies on the backbone but not the full model. It can also be used for MatchNet and ProtoNet.