Progress - Sprint 48

Done:

  • Fixed FFT and SNR differences
  • Custom scaled loss function
  • Custom output activation function for f_min (also tried for score, but doesn't work very well → sample weights?)

In progress

  • Pass updated predictions to Mike for usability test (especially in terms of f_min)

Todo

  • Gridsearch
  • Validation

  • de-skew?

Progress - Sprint 47

Done:

  • Active learning iteration f_min based

  • Custom output act function (+ sample weights) – (to discuss)
  • custom loss function
  • Added "bad" f_min records back in
  • SNR scaling
  • CNN for SNR input

In progress

  • Pass updated predictions to Mike

Todo:

  • Gridsearch
  • Validation

  • de-skew?

  • f_min improvements (increase frequency range of SNR values)

Progress - Sprint 46

Done:

  • Feature updates x2
  • Re-ran feature extraction for updated features
  • Switched to record based multi-output NN
  • Fixed non-convergence issue

In progress

  • Active Learning – about to pass Mike first set of new record ids to label (score-based)

  • Active learning iteration f_min based

Todo:

  • Gridsearch
  • Validation
  • de-skew?
  • sample weighting?
  • custom loss function?

Progress - Sprint 45

Done:

  • Existing functionality
  • Feature extraction for the full set of records
    • Solved multiple record issues & investigated differences between new & Xavier's dataset


In Progress

  • Active learning workflow
  • Extracting features per component (instead of geo mean)


Todo:

  • Gridsearch
  • de-skew?
  • Validation workflow

Progress - Sprint 44

Done:

  • Re-implementation of existing functionality
    • Feature extraction (Some issues, see below)
    • Run Canterbury and Canterbury-Wellington from feature csv and records


In Progress:

  • Data issues
  • Exploratory work and initial gauge initial subduction performance
    • Looked at deskew vs no-deskew
    • Only use best records vs all records (with and without weighting)
    • Train model on all shallow and see how it does on subduction --baseline

Future:

  • Compare trained with original on full validation dataset
  • Improved validation workflow
  • Gridsearch
  • Active learning for subduction

Data issues:

  • Some records have no buffer start date, only event start date, so should still be able to use these (currently throws an error)
  • Some records in the GMR.csv from Xavier are weird, not meeting x-axis crossing condition
  • malloc error, no idea?
  • 3744 records instead of 3989


  • No labels