inference_config.yml

# VeridisQuo Production Configuration - Complete parameter registry for inference engine deployment

type: inference

name: inference_config

paths:
  # File paths and URLs for models and resources
  project_root: ./
  models:
    deepfake_detection:
      dir: models/deepfake_detection
      filename: veridisquo_40M.pth
      auto_download: true
      huggingface:
        repo_id: Gazeux33/VeridisQuo
        model_filename: veridisquo_40M.pth
        url: https://huggingface.co/Gazeux33/VeridisQuo/resolve/main/veridisquo_40M.pth
    face_detection:
      custom_weights_path: models/face_detection/model.pt 
      enabled: true
    efficientnet:
      # EfficientNet weights - null uses ImageNet pretrained
      custom_weights_path: null
    frequency_extractor:
      # Custom weights for frequency feature extractor MLP fusion
      custom_weights_path: null
    classifier:
      # Custom weights for final classifier MLP
      custom_weights_path: null
  output:
    # Directory to save logs
    logs_dir: logs
    # Directory to save inference results
    results_dir: output

device:
  # Device configuration for inference
  # Options: auto, cpu, cuda, cuda:0, cuda:1
  # Description: auto selects CUDA if available, otherwise CPU
  type: auto

model_architecture:
  # DeepFakeDetector model architecture parameters

  # Input channel mode for images
  # Implemented channel mode options: luminance, grayscale, YCbCr, RGB
  channel_mode: luminance

  spatial_features:
    # EfficientNet-B4 for spatial feature extraction
    # Description: Pretrained on ImageNet, extracts spatial features from faces
    name: EfficientNet-B4
    output_dim: 1792

  frequency_features:
    # FFT and DCT combined frequency analysis
    # Description: Use PyTorch (GPU) vs NumPy (CPU) extractors
    use_gpu_extractors: true

    fft:
      # Fast Fourier Transform parameters
      feature_dim: 512
      num_radial_bands: 8
      window_function: hann
      high_freq_emphasis: true

      # FFT calculation constants
      constants:
        epsilon: 1e-8
        num_azimuthal_sectors: 8
        high_freq_threshold_ratio: 0.5
        artifact_num_radial_samples: 50
        artifact_center_region_size: 20
        artifact_edge_width: 5
        default_smoothing_kernel_size: 5
        energy_band_start_multiplier: 2
        energy_band_end_multiplier: 3

    dct:
      # Discrete Cosine Transform parameters
      feature_dim: 512
      block_size: 8
      aggregation_method: frequency_bands
      num_frequency_bands: 4

      # DCT calculation constants
      constants:
        high_freq_emphasis: true
        epsilon: 1e-8

    fusion:
      # MLP fusion of FFT and DCT features
      output_dim: 1024
      hidden_dims: null
      dropout_rate: 0.3
      use_batch_norm: true

  concatenation:
    # Concatenation of spatial and frequency features
    spatial_dim: 1792
    frequency_dim: 1024

  classifier:
    # Final MLP classifier for class prediction
    # Implemented classes: 2 (REAL, FAKE)
    input_dim: 2816
    num_classes: 2
    hidden_dims: null
    dropout_rate: 0.2
    use_batch_norm: true

preprocessing:
  # Image and video preprocessing parameters

  image:
    # Format: [width, height]
    # Description: All images resized to 224x224 and normalized using ImageNet stats to match EfficientNet input requirements
    target_size: [224, 224]
    normalization:
      # Description: ImageNet normalization statistics (RGB)
      mean: [0.485, 0.456, 0.406]
      std: [0.229, 0.224, 0.225]

  face_detection:
    # YOLO-based face detection for video inference
    # Description: Detects faces in frames, filters by size and confidence
    enabled: true
    min_face_size: 40
    confidence_threshold: 0.7
    only_keep_top: true

  face_extraction:
    # Face region extraction and preprocessing
    # Description: Extracts face region from frame with optional padding. Normalization handled by torchvision in InferenceEngine
    # Padding unit: pixels
    # Normalization options: zero_one, minus_one_one, null
    target_size: [224, 224]
    padding: 0
    # Custom normalization method for face crops - If null, uses ImageNet stats from image normalization
    normalization_method: null 

  frame_extraction:
    # Video frame sampling parameters
    # Description: Number of frames to extract per second of video
    # max_frames description: null means no limit, integer limits total frames analyzed
    # optimized description: Use PyAV-based GPU-accelerated extractor (5-10x faster) if available
    frames_per_second: 1
    max_frames: null
    # Two frame extractors: optimized (GPU, using PyAV) and standard (CPU)
    # If optimized is selected but not available, falls back to standard extractor
    use_optimized: true

inference:
  # Inference runtime specific parameters

  batch_processing:
    # Description: Batch multiple frames for GPU efficiency
    enabled: true
    default_batch_size: 32

  score_aggregation:
    # ScoreAggregator parameters to combine frame-level predictions into single video prediction
    # Implemented methods: majority, average, weighted_average, threshold
    aggregation_method: majority
    # Several methods use thresholds to determine final class according to probability
    threshold: 0.5

  output_format:
    # Inference result formatting
    # Description: Control verbosity of inference results
    include_frame_results: true
    include_metadata: true
    include_raw_logits: false

logging:
  # Logging configuration
  enabled: true
  level: DEBUG
  format: "%(asctime)s - %(name)s.%(funcName)s - %(levelname)s - %(message)s - %(lineno)d"
  datefmt: "%Y-%m-%d %H:%M:%S"
  file:
    enabled: true
    # Path relative to logs_dir
    path: inference.log
    backup_count: 5
  console:
    enabled: true
    level: DEBUG

performance:
  # Performance optimization settings

  gpu:
    enabled: auto
    # Description: Enable cuDNN benchmark for optimized performance on fixed-size inputs
    cudnn_benchmark: true

  threading:
    # CPU threading for NumPy operations
    num_threads: null

  caching:
    # Model and preprocessing caching
    # Description: Trade memory for speed by caching intermediate results
    cache_preprocessed_frames: false
    cache_extracted_features: false

production:
  # Production deployment specific settings
  api:
    # Description: REST API configuration for production deployment
    enabled: false
    host: 0.0.0.0
    port: 8000
    workers: 4