Swarm description for hot gym anomaly example?

I’m wondering if it would be possible to make the swarm description for the one hot gym anomaly tutorial available on github? When I use model parameters which I have produced myself, doing my own swarming, I’m unable to get the nice anomaly scores and anomaly likelihoods which result if I use the model parameters in model_params/rec_center_hourly_model_params.py.
When I run swarming I used swarm_description.py from the prediction tutorial, except that I changed “TemporalMultiStep” to “TemporalAnomaly”. I also set “iterationCount” to -1 and “swarmSize” to “large”.

1 Like

There was no swarm for the hot gym anomaly example. I just re-used the model params from the prediction example. The first few minutes of this video explains it:

Thank you Matt for your quick reply! My response was less quick since I squeezed in a large swarming in between. Finally I succeeded with the anomaly scores, and I learned something: the swarming produces a better model if I set “inferenceType”: “TemporalMultiStep”, than if I set “inferenceType”: “TemporalAnomaly” (in swarm_description.py). Isn’t that a bit odd? After all I used the model for anomaly calculation.

The swarming process is only looking for parameters that improve the quality of predictions on the predictedField. There is no way to swarm to optimize for anomaly detection. I am surprised the swarm even worked with inferenceType = TemporalAnomaly!