search

aai-institute / ProtoTree

Fork of ProtoTrees: Neural Prototype Trees for Interpretable Fine-grained Image Recognition, published at CVPR2021
Other
3 stars 1 forks source link

Take the minimum Euclidean distances for prototypes instead of the maxima #10

Closed Clifford-appliedAI-GmbH closed 1 year ago

Clifford-appliedAI-GmbH commented 1 year ago

Fixes the bug in which the tree was taking the maximum distances between latent patches and prototypes instead of the minima. With this fix the model goes from training loss around 4 and test accuracy around 10%, to 0.5 loss and 88% accuracy.