config.example.yaml

# MuSc Industrial GUI Configuration Example
# Copy this file to config.yaml and modify as needed

datasets:
  class_name: ALL  # Category to test, or ALL for all categories
  data_path: ./data/  # Path to your dataset (relative to project root)
  dataset_name: mvtec_ad  # Dataset name (mvtec_ad, visa, btad)
  divide_num: 1  # Number of subsets to divide test set into
  img_resize: 224  # Image resize dimension (224, 256, 384, 512)

device: 0  # GPU device index (0, 1, etc.) or 'cpu' for CPU-only mode

models:
  # Backbone model - choose from:
  # CLIP models: ViT-B-32, ViT-B-16, ViT-L-14
  # DINO models: dino_deitsmall16, dino_vitbase16, dino_vitbase8
  # DINOv2 models: dinov2_vitb14, dinov2_vitl14
  # Other: google/siglip-so400m-patch14-384, vit_small_patch32_224.augreg_in21k,
  #        vit_tiny_patch16_224.augreg_in21k, vit_small_patch16_224.dino
  backbone_name: dinov2_vitb14

  batch_size: 1  # Batch size for inference

  feature_layers:
  - 11  # Layer indices to extract features from

  pretrained: openai  # Pretrained model source (openai, laion400m_e31, laion400m_e32)

  r_list:
  - 1  # Aggregation degrees for LNAMD module

testing:
  output_dir: output_gui  # Directory for saving outputs
  save_excel: false  # Whether to save results to Excel
  vis: false  # Whether to save visualization images
  vis_type: single_norm  # Visualization normalization type (single_norm or whole_norm)

thresholds:
  # Image threshold: 1.0-10.0 in GUI (divided by 10 internally to get 0.1-1.0)
  # Higher = fewer false positives, may miss subtle anomalies
  # Lower = more sensitive, more false positives
  # Recommended: 8.0-9.5 for most applications
  image_threshold: 9.0

  # Overlay threshold: 1.0-10.0 in GUI (divided by 10 internally to get 0.1-1.0)
  # Controls visualization overlay intensity
  # Does not affect detection, only visualization
  # Recommended: 3.0-5.0 for clear visualization
  overlay_threshold: 3.5