""" Global configuration for the classical music grid Transformer. """ # ========================= # Grid and tempo constants # ========================= # Time grid STEPS_PER_BAR: int = 8 INPUT_BARS: int = 3 INPUT_STEPS: int = INPUT_BARS * STEPS_PER_BAR # 24 # Pitch grid PITCHES_MELODIC: int = 12 # one octave: 12 pitch classes PITCHES_PERC: int = 3 # e.g., Kick, Snare, Hi-hat # Instrument families (for clarity / future use) N_FAMILIES: int = 4 # Strings, Woodwinds, Brass, Percussion # Tempo and meter TEMPO_BPM: int = 120 BEATS_PER_BAR: int = 4 # assuming 4/4 time # Derived timing values SECONDS_PER_BEAT: float = 60.0 / TEMPO_BPM SECONDS_PER_BAR: float = BEATS_PER_BAR * SECONDS_PER_BEAT SECONDS_PER_STEP: float = SECONDS_PER_BAR / STEPS_PER_BAR # ========================= # Model hyperparameters # ========================= # Transformer dimensions D_MODEL: int = 256 # hidden size N_LAYERS: int = 3 # number of Transformer encoder layers (2–4 is reasonable) N_HEADS: int = 4 # attention heads FFN_DIM: int = 512 # feed-forward inner dimension DROPOUT: float = 0.1 # dropout rate # Input feature size per time step: # Strings (12) + Woodwinds (12) + Brass (12) + Percussion (3) = 39 INPUT_FEATURES: int = (3 * PITCHES_MELODIC) + PITCHES_PERC # 39 # ========================= # Sampling and constraints # ========================= # Bar-level similarity threshold (0–1); above this, bar is considered "too similar" BAR_SIMILARITY_THRESHOLD: float = 0.9 # Maximum number of resampling attempts per bar before accepting anyway MAX_RESAMPLE_ATTEMPTS_PER_BAR: int = 5 # Minimum number of instrument families that must be active (have ≥1 note) per bar MIN_ACTIVE_FAMILIES_PER_BAR: int = 2