schatty / prototypical-networks-tf Goto Github PK

When I train on my custom dataset. Acc on training data is 95% but when I use backbone CNN to extract image into embedding vector, then I compare distance with cluster, the acc is so low? Is there something wrong with me? Or Should I use different approaches?
Thanks!

After training and saving the model, how do I use the weights file to predict the class of a new unlabelled image?
Do I always have to feed in a support set during prediction?