eqModelTransformer.yaml

seml:
  executable: eqModelTrain.py
  name: eqModelTrain
  output_dir: logs
  project_root_dir: .


slurm:
  experiments_per_job: 5
  max_simultaneous_jobs: 6  # Restrict number of simultaneously running jobs per job array
  sbatch_options:
    gres: gpu:1       # num GPUs
    mem: 35G          # memory
    cpus-per-task: 96  # num cores
    time: 1-08:00     # max time, D-HH:MM

fixed:

  init.model_chosen: 'main' # ['main,'nofeatures']
  init.seed_value: 1 # seed value for everything (still something makes the run not deterministic atm.)
  init.patience: 15 # patience while running

  data.doFT: True # do the SFA as preprocessing
  
  model.useEmbed: False # currently always false and a mapping is used as embedding
  model.useSaves: False # use already trained models rather than train again


  model.steps: ['max'] # Which steps to try out for LASA. ['max', 'sum'] #only those two are supported
  model.layerCombis: ['hl'] # ['lh', 'hl'] as alternative for LASA combination steps
  model.localAvgThresholds: [[0.5,0.5],[1.0,1.0],[1.4,1.4]] # thresholds for lasa

  model.header: 6 # number of transformer heads

  model.numOfAttentionLayers: 2 # number of transformer layers

  model.dffFaktor: 0.5 # dff size in ralation to dmodel, e.g. if dmodel = 16 and dffFaktor = 0.5, dff is 8

  model.transDropout: 0.0 # transformer dropout. Not suggested.

  model.lastDropout: 0.4 # dropout at the end of the model

  model.gcnNodes: 32 # number of gcn nodes

  model.finalLayer: 256 # final dense layer size
  model.filters: [16,32] # number of filters for the cnn part
  model.cnnKernal: 25 # kernal size for the cnn part (make sure it fits the ncoef size)

  model.modelN: 'transformer' # Current file is optimized for the transformer model. Alternatives: 'old','bloem' 
  model.doAbstract: True # do lasa after training (only for transformer)
  model.doMixedLasa: False # switch dimensions and apply lasa again

  model.epochs: 200 # number of max epochs

  

grid:

  data.network_choice: # which of both networks to train
      type: choice
      options: 
        - network1
        - network2

  

  model.doStations: # if doStations is false the model uses the following shape: (time points, features * stations) else (stations, features * time points). This is important for lasa in the sense of what dimension to reduce. Only works with the transformer
      type: choice
      options:
        - True
        - False

  data.random_state_here: # state for data splits
    type: choice
    options: 
      - 1
      - 2
      - 3
      - 4
      - 5

  data.nbins: # number of symbols used for sfa
    type: choice
    options: 
      - 5
      - 6
      - 7

  data.ncoef: # number of coefs used from the SFA. This also is afterwards the size of the last dim times the number of features
    type: choice
    options: 
      - 125
      - 250

  data.iLen: # length of the time window used for the prediction. 1000 = 10s
    type: choice
    options:
      - 1000
      - 900
      - 800
      - 700
      - 600
      - 500
      - 400