Different number of detected events for multiple runs

[smousavi05]

Different number of detected events for multiple runs

[smousavi05]

Hi @hycarbon this is totally normal and due to applying dropout sampling during each prediction. The output probabilities differs in each inference and this is what we basically use to estimate the uncertainties in the model estimates. Different number of detected events you get is simply for those events with low probability (or high uncertainties) that running the model for a second time result output probability bellow the threshold value. If you set the university estimate option on in the input parameters the code will do multiple predictions basically and the resulted output would be the average of them.

[YukuanChen]

Hi @smousavi05 thanks for your quick reply.

I'm using mseed_predictor, it seems no multiple prediction option in mseed_predictor and it can not estimate uncertainties, so I would like to check again with you that if there still are dropout sampling in mseed_predictor. (I know the predictor working on hdf5 files has this option and can estimate uncertainties of results).

Actually, I found some clear P and S signals not detected by EQT even though these signals are very similar to those detected. I'm wondering if this is caused by the randomness in EQT, do you have any suggestions? Thanks.

[smousavi05]

@hycarbon Both mseed_predictor and predictor modules use the same model that includes the dropout sampling. The can be multiple reasons for those missed picks. Dropout sampling can have an effect but sometimes the stations characteristics and out of distribution problem can be the case. However there shouldn’t be too many false negatives. There are multiple solutions. The first one is transfer learning which might be a bit time consuming as include some training and fine tuning. Another approach is to use the detected events by EqT as templates and perform a template matching to make sure nothing has left out. Our experiments should this might increase the number of events by 10 %. There are some ml tricks as well that might help but those are more complicated.

[YukuanChen]

Thanks.