config.yaml

# =============================================================================
# Configuration File for Training the Transformer Model
#
# You can rename config parameters, add or remove sections as you wish.
# The script will read these fields by their keys.
#
# NOTE: YAML requires correct indentation. Comments are prefixed by '#'.
# =============================================================================

paths:
  # input_csv: Path to your CSV file containing data.
  #            This can be changed for different CSVs to retrain on new data.
  #
  input_csv: "falls_detected.csv"
  # input_csv: "fallD_batt_empty.csv"
  # output_plots_dir: Directory to save any output plots (optional usage).
  #                   If you do not wish to save plots, you can ignore or remove it.
  output_plots_dir: "plots"

columns:
  # date_col: Name of the date/time column used to parse dates (must exist in your CSV).
  date_col: "Date"

  # target_col: The column name in your CSV that you want to forecast.
  # target_col: "Battery_Empty"
  target_col: "Falls"
  # feature_cols: The list of column names (as they appear in CSV) to be used as features.
  #               You can include the target_col itself here if you also use it as an input feature.
  #               This script automatically adds day-of-week features, but you can rename or remove them if needed.
  feature_cols:
    - "Falls"
    #- "Battery_Empty"
    - "dow_sin"
    - "dow_cos"
    - "is_weekend"

training:
  # window_size: Number of past time steps to include in a single input sequence for the model.
  #              A larger window_size can help the model see more history (but can also increase training complexity).
  window_size: 30

  # train_split: Fraction of data used for training (remainder is for validation/testing).
  #              e.g. 0.8 means 80% training, 20% testing.
  train_split: 0.8

  # epochs: Number of epochs (full passes through the training data).
  #         Higher epochs can improve results but risk overfitting if too large.
  epochs: 20

  # batch_size: Number of samples per gradient update.
  #             Smaller batches can lead to more fine-grained updates, but slower training.
  batch_size: 16

  # learning_rate: The step size for the optimizer when updating weights.
  #                Too large => might not converge, too small => slow training.
  learning_rate: 0.000001

  # 26, 16 for falls detected, good predictions
  # learning_rate: 0.000001


transformer:
  # d_model: Dimensionality for the Transformer encoding space.
  #          Must be large enough to capture information from your input features.
  d_model: 32

  # num_heads: Number of attention heads in MultiHeadAttention.
  #            More heads can capture more diverse relationships in time series, but also more computational cost.
  num_heads: 8

  # ff_dim: Dimensionality of the feed-forward network within each Transformer block.
  #         Typically larger than d_model to allow complex transformations.
  ff_dim: 128

  # num_transformer_blocks: How many times to stack the Transformer encoder block.
  #                         Deeper stacks => more representational power but can overfit if your dataset is small.
  num_transformer_blocks: 4

  # dropout_rate: Dropout fraction in the attention and dense layers. Helps prevent overfitting.
  #               Usually between 0.0 and 0.5 for typical deep learning tasks.
  dropout_rate: 0.01

    #  EXAMPLES FOR smooth

    # d_model=32,
    # num_heads=2,
    # ff_dim=128,
    # num_transformer_blocks=2,
    # dropout_rate=0.01

    #  EXAMPLES FOR FALLS DETECTED 1 GREAT
    # EPOCH 20,16
    # d_model=32
    # num_heads=2, also 8
    # ff_dim=128,
    # num_transformer_blocks=2, also 4
    # dropout_rate=0.02

    #  WORKS FOR FALLS DETECTED
    # 30 epoch with 16 size
    # d_model=128,
    # num_heads=4,
    # ff_dim=512,
    # num_transformer_blocks=4,
    # dropout_rate=0.01

forecast:
  # n_future: How many future time steps to forecast iteratively.
  n_future: 15