I’ve officially discontinued the Earthquake Predictor project for now. I still need to do backups of the data, but I’m no longer going to have the daily update script running. It had stopped for a while now due to a bug anyway, and I wasn’t able after a cursory effort to debug it so that it worked again. It also honestly didn’t do anything relevant, as the model only predicted high frequency, low magnitude quakes and not the low frequency, high magnitude ones that mattered. The architecture I was using was my old LSTM-RNN model, which has largely been superseded in the literature by the transformer architecture, so it badly needed to be taken offline and retrained anyway. Not sure if I’ll ever get around to retraining it with the new architecture. It kinda seems like a very silly longshot of a project that kept one entire GPU busy because I could never figure out how to properly unload models from memory without stopping the process. Even if it only updated once a day, it was a hassle because the thing would randomly error out and I’d have to start it up again after having missed several days of updates and having to backdate the sequence of updates made with the appropriate 24 hours of data. All in all, not a great system.
