search

ninatu / anomaly_detection

This is the official implementation of "Anomaly Detection with Deep Perceptual Autoencoders".
Apache License 2.0
44 stars 8 forks source link

Validation set for the early stopping #4

Open cuicathy opened 1 year ago

cuicathy commented 1 year ago

Hi,

Thanks for your great work! I have a question about the early stopping. According to your paper, 200 tumor patches (abnormal samples) were used as the validation set. In your code, I noticed that the sum of the adversarial loss and reconstruction loss of the validation set was used as the sign for early stopping. I am a little confused why the loss score of abnormal samples in the validation set instead of the AUC score of validation sets with both abnormal and normal images available is used for early stopping.

I am looking forward to your reply. Thank you!

ninatu commented 1 year ago

Hi!

Thanks for the question. We used loss on the validation set for earlier stopping to avoid overfitting on the train data (with the given set hyperparams). And then, we use ROC AUC to choose the best hyperparameters. However, you are right, ROC AUC may also be used for earlier stopping.