Skip to main content

Configuration Guide

Overview​

SpikeYoloV8-Tracker uses a centralized configuration system through config/config.yaml. All project settings are managed in one place, with no hardcoded values.

Configuration File Structure​

The main configuration file is located at config/config.yaml. It contains sections for:

  • Model Configuration: Architecture and model parameters
  • Training Configuration: Training hyperparameters and settings
  • Data Processing: Data loading and preprocessing settings
  • Loss Configuration: Loss function weights and parameters
  • Logging Configuration: Logging settings and paths

Key Configuration Sections​

Model Configuration​

model:
  name: "SpikeYOLO_Tracker"
  num_classes: 8  # Automatically detected from classes list
  classes:
    - Pedestrian
    - Car
    - Bicycle
    - Bus
    - Motorbike
    - Truck
    - Tram
    - Wheelchair
  logs_dir: "./logs"

Training Configuration​

training:
  epochs: 25
  warmup_epochs: 3
  batch_size: 25
  learning_rate: 0.001
  optimizer: "sgd"  # Options: "sgd", "adamw"
  momentum: 0.9
  lr_scheduler: "step"  # Options: "step", "cyclic"
  max_learning_rate: 0.002  # For cyclic scheduler
  device: "cuda"  # Options: "cuda", "cpu"

Data Processing Configuration​

data_processing:
  max_samples_per_file: 50
  max_events_per_sample: 10000
  temporal_buffer: 0.2  # 20% buffer for temporal matching
  slice_duration_us: 100  # Time window in microseconds
  time_steps: 8  # Number of temporal steps
  prefetch_factor: 2
  persistent_workers: true
  pin_memory: true

Loss Configuration​

yolo_loss:
  box_loss_weight: 5.0  # Increased from 2.0 for better localization
  class_loss_weight: 1.0
  obj_loss_weight: 1.0
  track_loss_weight: 1.0
  use_focal_loss: true
  focal_alpha: 0.25
  focal_gamma: 2.0
  label_smoothing: 0.1  # Prevents overconfidence

Configuration Access​

The project uses src/config_loader.py for configuration management:

from src.config_loader import ConfigLoader

# Load configuration
config = ConfigLoader('config/config.yaml')

# Access configuration values
num_classes = config.get('model.num_classes')
batch_size = config.get('training.batch_size')
learning_rate = config.get('training.learning_rate')

Dynamic Class Configuration​

Classes are defined dynamically in the configuration file. The number of classes is automatically detected:

model:
  classes:
    - Pedestrian
    - Car
    - Bicycle
    # Add more classes as needed

The system automatically:

  • Detects the number of classes from the list
  • Updates model architecture accordingly
  • Handles class mappings in annotations

Training Optimizations Configuration​

Warmup Learning Rate​

training:
  warmup_epochs: 3  # Gradual LR increase prevents cold start

Cyclic Learning Rate​

training:
  lr_scheduler: "cyclic"
  max_learning_rate: 0.002
  # OneCycleLR scheduler helps escape poor local minima

Focal Loss​

yolo_loss:
  use_focal_loss: true
  focal_alpha: 0.25
  focal_gamma: 2.0
  # Focuses learning on hard-to-classify examples

Label Smoothing​

yolo_loss:
  label_smoothing: 0.1
  # Softens hard labels to prevent overconfidence

Data Loader Configuration​

data_processing:
  prefetch_factor: 2  # Number of batches to prefetch
  persistent_workers: true  # Keep workers alive between epochs
  pin_memory: true  # Faster GPU transfer

Logging Configuration​

model:
  logs_dir: "./logs"
  # Logs are automatically created in this directory:
  # - training_{timestamp}.log
  # - evaluation_{checkpoint_name}.log
  # - hyperparameter_search/hyperparameter_search_{trial_name}.log

Configuration Best Practices​

  1. Centralized Management: All settings in config.yaml
  2. No Hardcoded Values: Use config loader for all parameters
  3. Dynamic Classes: Define classes in config, not in code
  4. Version Control: Keep config files in version control
  5. Environment-Specific: Use different configs for dev/prod if needed

Example: Complete Configuration​

model:
  name: "SpikeYOLO_Tracker"
  classes:
    - Pedestrian
    - Car
    - Bicycle
    - Bus
    - Motorbike
    - Truck
    - Tram
    - Wheelchair
  logs_dir: "./logs"

training:
  epochs: 25
  warmup_epochs: 3
  batch_size: 25
  learning_rate: 0.001
  optimizer: "sgd"
  momentum: 0.9
  lr_scheduler: "step"
  device: "cuda"

data_processing:
  max_samples_per_file: 50
  max_events_per_sample: 10000
  temporal_buffer: 0.2
  slice_duration_us: 100
  time_steps: 8
  prefetch_factor: 2
  persistent_workers: true
  pin_memory: true

yolo_loss:
  box_loss_weight: 5.0
  class_loss_weight: 1.0
  obj_loss_weight: 1.0
  track_loss_weight: 1.0
  use_focal_loss: true
  focal_alpha: 0.25
  focal_gamma: 2.0
  label_smoothing: 0.1

Next Steps​