Hi, I call swarm to train my data mode, then use the model_params.py which generated via swarm to run NuPIC, but the anomalyScore is none. Is anything I misunderstand? Thanks.
Swarming does not generate anomaly models. You can change this in the generated model params by changing the inferenceType
to TemporalAnomaly
, I think.
For more details about changing a prediction model into an anomaly model, see my anomaly tutorial.
Thank you for your explanation. I changed inferenceType to "TemporalAnomaly"and re-swarm, then I can get the anomalyscore. But I am confused the result of likelihood score, looks like the first a couple of scores are the same score Test_Output.csv. Is it normal? How many values will skip for anomaly detection? I mean how many scores are the same at the beginning of input.
No, that’s not what I meant. After the swarm, you take the params it gives you and change it to an anomaly model. There is no need to re-swarm. Also, I forgot that @scott and I talked about this last year on youtube.
Regarding anomaly likelihood, because it is a windowed operation, it needs to see a certain number of records before it will start producing meaningful values. I would not trust anything it outputs until 500 rows of data have been processed.
It doesn’t work like that. Here are docs about how the anomaly score is calculated from internal neuron states. There is also a detailed video on the subject: