params.yaml

# Angle Grinder Detection - Configuration Parameters
# Edit these values to customize your pipeline

# ============================================================================
# DATA AUGMENTATION
# ============================================================================
augmentation:
  grinder_multiplier: 11         # Multiply grinder samples by 25 (~36 → ~900)
  tools_multiplier: 9             # Multiply tool samples by 20
  background_multiplier: 1         # No augmentation for background (already diverse)
  
  # Noise addition (SNR in dB)
  noise_snr_min: 12                # Minimum signal-to-noise ratio
  noise_snr_max: 18                # Maximum signal-to-noise ratio
  
  # Time stretching (playback speed while maintaining pitch)
  time_stretch_min: 0.85           # 0.85x speed (slower)
  time_stretch_max: 1.15           # 1.15x speed (faster)
  
  # Pitch shifting (semitones)
  pitch_shift_min: -3              # Down 3 semitones
  pitch_shift_max: 3               # Up 3 semitones
  
  # Gain adjustment (dB)
  gain_min: -4                     # Quieter by 4 dB
  gain_max: 4                      # Louder by 4 dB

# ============================================================================
# PREPROCESSING
# ============================================================================
preprocessing:
  target_sr: 16000                 # Changed from 16000
  segment_duration: 1.0            # Keep at 1 second
  segment_overlap: 0.5             # Keep at 50%
  hop_length: 8000                # Changed from 8000 (0.5s at 44.1kHz)
  window_size: 16000              # Changed from 16000 (1s at 44.1kHz)

  
  # High-pass filter (remove low frequency rumble)
  highpass_cutoff: 80              # Cutoff frequency in Hz
  highpass_order: 5                # Filter order (steepness)
  
  # Normalization
  norm_method: 'peak'              # Normalization method ('peak' or 'rms')
  peak_level: 0.89                 # Target peak level (-1 dB to avoid clipping)

# ============================================================================
# FEATURE EXTRACTION - CLASSICAL ML
# ============================================================================
features_classical:
  # MFCC (Mel-Frequency Cepstral Coefficients)
  n_mfcc: 13                       # Number of MFCC coefficients
  n_fft: 512                       # FFT window size
  hop_length: 160                  # Samples between frames (10ms at 16kHz)
  
  # Spectral features
  spectral_rolloff_percent: 0.85   # Frequency below which 85% of energy is contained
  spectral_contrast_bands: 6       # Number of frequency bands (produces 7 bands total)
  
  # GFCC (Gammatone Frequency Cepstral Coefficients) using spafe
  gfcc_n_filters: 100              # Number of gammatone filters
  gfcc_fmin: 50                    # Minimum frequency
  gfcc_fmax: 8000                  # Maximum frequency
  gfcc_n_coeffs: 13                # Number of GFCC coefficients
  
  # Feature sets to extract
  feature_sets:
    - mfcc                         # MFCC-only baseline (~194 features)
    - gfcc                         # GFCC-only (~194 features)
    - combined                     # All features combined (~300 features)

# ============================================================================
# FEATURE EXTRACTION - NEURAL NETWORKS
# ============================================================================
features_neural:
  n_fft: 512                       # FFT window size
  hop_length: 160                  # Samples between frames
  n_mels: 40                       # Number of mel frequency bands
  n_channels: 3                    # 1: log-mel only, 3: log-mel + delta + delta-delta

# ============================================================================
# DATA SPLITTING
# ============================================================================
data_split:
  train_ratio: 0.70                # 70% for training
  val_ratio: 0.15                  # 15% for validation
  test_ratio: 0.15                 # 15% for testing
  stratify: true                   # Maintain class balance across splits
  random_state: 42                 # Random seed for reproducibility
  split_by_file: true              # Split at file level to prevent data leakage

# ============================================================================
# CLASS BALANCING
# ============================================================================
balancing:
  method: 'hybrid'                 # 'hybrid', 'oversample', 'undersample', 'none'
  undersample_ratio: 0.8           # Reduce non-grinder to 0.8:1 ratio before ADASYN
  oversample_ratio: 1.0            # ADASYN to achieve 1:1 final ratio
  apply_to: 'train'                # Only balance training set

# ============================================================================
# CLASSICAL ML - HYPERPARAMETER SEARCH SPACES
# ============================================================================
classical_ml:
  # Support Vector Machine
  svm:
    C: [0.1, 1, 10, 100]           # Regularization parameter
    gamma: ['scale', 'auto', 0.001, 0.01]  # Kernel coefficient
    kernel: ['rbf']                # Kernel type
  
  # Random Forest
  random_forest:
    n_estimators: [50, 100, 200]   # Number of trees
    max_depth: [10, 20, 30, null]  # Maximum tree depth (null = no limit)
    min_samples_split: [2, 5, 10]  # Minimum samples to split node
    max_features: ['sqrt', 'log2'] # Features to consider per split
  
  # XGBoost
  xgboost:
    learning_rate: [0.01, 0.05, 0.1]         # Step size shrinkage
    max_depth: [3, 5, 7, 9]                  # Maximum tree depth
    n_estimators: [100, 200, 300]            # Number of boosting rounds
    min_child_weight: [1, 3, 5]              # Minimum sum of weights in child
    subsample: [0.8, 0.9, 1.0]               # Subsample ratio of training data
    colsample_bytree: [0.8, 0.9, 1.0]        # Subsample ratio of features
  
  # K-Nearest Neighbors
  knn:
    n_neighbors: [3, 5, 7, 11, 15, 21]       # Number of neighbors
    weights: ['uniform', 'distance']         # Weight function
    metric: ['euclidean', 'manhattan']       # Distance metric
  
  # Gaussian Mixture Model
  gmm:
    n_components: [2, 4, 6, 8]               # Number of mixture components
    covariance_type: ['full', 'diag']        # Covariance matrix type

# ============================================================================
# NEURAL NETWORKS
# ============================================================================
neural:
  # Custom CNN
  cnn:
    batch_size: 64                 # Samples per batch
    epochs: 30                     # Maximum training epochs
    learning_rate: 0.001           # Initial learning rate
    dropout_rate: 0.5              # Dropout for regularization
    early_stopping_patience: 5     # Epochs without improvement before stopping
    reduce_lr_patience: 3          # Epochs before reducing learning rate
    reduce_lr_factor: 0.5          # Factor to reduce learning rate
  
  # YAMNet Transfer Learning
  yamnet:
    batch_size: 128                # Larger batch size for transfer learning
    epochs: 15                     # Fewer epochs needed
    learning_rate: 0.001           # Initial learning rate for head
    head_units: [128, 64]          # Dense layer sizes in classification head
    dropout_rates: [0.4, 0.3]      # Dropout for each dense layer
    early_stopping_patience: 3     # Fewer epochs needed
    
    # Fine-tuning settings
    finetune_learning_rate: 0.0001 # 10x slower for fine-tuning
    finetune_layers: 4             # Number of YAMNet layers to unfreeze
  
  # Knowledge Distillation
  distillation:
    temperature: 4                 # Softening factor for teacher predictions
    alpha: 0.7                     # Weight for soft loss (vs hard loss)
    epochs: 20                     # Distillation training epochs

# ============================================================================
# OPTIMIZATION
# ============================================================================
optimization:
  n_selected_features: 62          # Number of features after selection
  imputation_estimators: 50        # Trees in imputation forest
  imputation_max_depth: 10         # Max depth for imputation trees
  
  # Reduced spectrogram resolution (for deployment)
  reduced_spec:
    n_fft: 256                     # Smaller FFT window
    hop_length: 320                # Larger hop (fewer frames)
    n_mels: 20                     # Fewer mel bands

# ============================================================================
# MLFLOW
# ============================================================================
mlflow:
  experiment_name: 'angle_grinder_pipeline'  # MLflow experiment name
  tracking_uri: './mlruns'                   # Local tracking directory
  autolog: true                              # Auto-log parameters and metrics

# ============================================================================
# DVC
# ============================================================================
dvc:
  remote: 'local_storage'          # Default remote name
  autostage: true                  # Automatically stage DVC files with Git

# ============================================================================
# EVALUATION
# ============================================================================
evaluation:
  metrics: ['accuracy', 'precision', 'recall', 'f1', 'roc_auc']  # Metrics to compute
  cv_folds: 5                      # Number of cross-validation folds
  conf_matrix: true                # Generate confusion matrices
  roc_curve: true                  # Generate ROC curves
  
# ============================================================================
# PATHS (relative to project root)
# ============================================================================
paths:
  data_raw: 'data/raw'
  data_processed: 'data/processed'
  data_features_classical: 'data/features/classical'
  data_features_neural: 'data/features/neural'
  models_classical: 'models/classical'
  models_neural: 'models/neural'
  models_quantized: 'models/quantized'
  results: 'results'
  logs: 'logs'