How to implement training with a large number of positive sample data sets?

[lemonxiaohei]

hello I have a large number of positive sample data sets, and the training finds that each epoch has to convert all the pictures in the training set into "feature" for sample, which leads to insufficient video memory. How do I solve this problem

[Sj-Yuan]

The patchcore is a memory-bank based method, which requires memorizing all features from training samples, it is computationally expensive. Therefore, if you have a large number of normal data sets, it is better to choose other method for training, such as reverse distillation, effcientAD.