config.yml

# Train command
TRAIN_COMMAND: [ python, upstream.py ]

# Path to training dataset 
TRAIN_DATA: /path/to/upstream_dataset/
# Amount of training data used for validation (between 0 and 1)
#VALID_SPLIT: 0.1
SEED: 256

NUM_WORKERS: 2  # Num of workers for data loading
NUM_GPUS: [0]  # How many GPUs to use
WANDB: False

PROJ_NAME: LTA2VUpstream  # Just for logging

# -- Dataset class
DATASET: HUNT4Masked
DATASET_ARGS:
  x_columns: [[0,1,2,3,4,5]]  # Which columns to use 0=Bx, 1=By, 2=Bz, 3=Tx, 4=Ty, 5=Tz
  y_columns: [[0,1,2,3,4,5]]
  switch_sensor_prob: [0.0]  # Switch x and y columns randomly with prob
  sample_freq: [50]  # Original sampling freq in Hz. Do not change
  sequence_length: [5_400_000]  # Window size (in samples) fed into the model, here 30hours
  phase: [False]
  normalize: [True]  # Whether to normalize the training data
  hour_wise_norm: [False]  # Consider hour-based weighted normalization (default=True)
  STFT: [True]  # Compute STFT before feeding into model
  stored_as_STFT: [True]  # If dataset is already stored as spectrograms
  stft_n_fft: [3_000]  # Size of Fourier transform (in samples), here 1min
  stft_hop_len: [null]  # For STFT computation. If None, n_fft//2
  num_train_hours: [657_360]  # Amount of HUNT4 hours (80% of dataset)
  num_valid_hours: [82_170]  # Amount of HUNT4 hours (10% of dataset)
  num_test_hours: [82_170]  # Amount of HUNT4 hours (10% of dataset)
  # Returns timestamps of x, i.e.: (x, x_ts), (y, mask)
  # Timestamps are 5-dim int vectors of the form: [month, day, weekday, hour, minute]
  return_timestamps: True
  # Whether to extract day time only (6am-12am)
  skip_night_time: False
  ##### Alteration parameters #####
  # percentage of time frames to mask in spectrogram:
  time_alter_percentage: 0.10
  # number of consecutive time frames to mask in spectrogram:
  time_alter_width: 5
  # swap time windows (like in Liu et al. 2021) (default=True)
  time_swap: False
  # Probability of flipping time mask (default is 0.0)
  time_mask_flip_prob: 0.5
  # percentage of frequency bins to mask in spectrogram:
  freq_alter_percentage: 0.05
  # number of consecutive frequency bins to mask in spectrogram:
  freq_alter_width: 75  # Around 5% of 1 spectrogram's freq bins
  #################################

# -- Model 
# Which classifier to use 
ALGORITHM: TransformerEncoderNetwork
# Arguments for classifier
# (all given as lists in case to perform a GridSearch)
ALGORITHM_ARGS:
  epochs: [15]
  batch_size: [4]
  loss: [maskedL1]
  output_activation: [null]  # null=identity
  metrics: [null]  # Which metrics to log
  optimizer: [AdamW]
  weight_decay: [1e-5]
  lr: [1e-4]
  lr_scheduler: [LinearWUSchedule]
  # Architecture params
  n_encoder_layers: [4]
  nhead: [6]
  d_model: [1500]
  dim_feedforward: [2048]  # Dim of ff layer inside encoder layer
  # Local and Global Positional Encoding:
  positional_encoding: [['TemporalPositionalEncoding', 'AbsolutePositionalEncoding']]
  dropout: [0.1]  # Dropout probability
  n_prediction_head_layers: [1]  # Num layers of MLP at end of model
  dim_prediction_head: [null]  # Dim of hidden layers of MLP at end od model
  # What opertation to perform on sequence output?
  # null="seq2seq model", mean="seq2single by averaging output sequences"
  seq_operation: [null]