Results

[bmr070]

Since the anomaly-transformer uses almost similar solver code I was able to bring in the model. On the NIPS TS SWAN dataset using the same parameters as DCdetector I was able to get a comparable F1 score. ~71 F1 for Anomaly Transformer vs ~73 for DCdetector. Which is much better than what you have listed for Anomaly-Transformer in your result4.jpg image.

[yyysjz1997]

Since the anomaly-transformer uses almost similar solver code I was able to bring in the model. On the NIPS TS SWAN dataset using the same parameters as DCdetector I was able to get a comparable F1 score. ~71 F1 for Anomaly Transformer vs ~73 for DCdetector. Which is much better than what you have listed for Anomaly-Transformer in your result4.jpg image.

Thanks for your query. I wonder if you tried adjusting the hyperparameters. I just ran it quickly on my other server, and it matched the result on paper.