Hey @lis07128 , as we discussed together here some todos for validating our anomaly detection: - [x] one loss value per window - [x] create ROC-curve (on batch level) - [x] set labels globally on False and label some anomaly windows (qualitiative identification) on True - [x] for qualitiative identification create helper fn "labeling_bias" to synchronise traj and better isolate anomalous windows - [x] choose values for the threshold that represent the whole spectrum (~ some FP and no TP) - [x] create confusion matrix (on batch level) - [ ] adapt the labels of the conf matrix that we have anomaly as first entry and nominal as second, leading to TP on 1 row -1 col and FP on 1-2, FN on 2-1 and TN on 2-2 - [x] nice would be to summarize the results on some slides