config.yaml

common:
  num_iters: 4 # Number of iterations to run pipeline
  final_stage: 'refine' # Final stage to execute in final iteration. Choose from 'classifier','generate_features', 'anomaly_detection','attribution', 'clustering','merge','refine'
  save_path: 'results/run_default' # Root path for saving results from each stage and iteration
  resume: True # Resume from previous completed stage
  train_data_path: 'data/train/' # Root path containing training images
  eval_data_path: 'data/eval/' # Root path containing evaluation images
  num_workers: 16 # Number of workers for dataloader
  eval_batch_size: 1024 # Batch size for feature generation eval

step1: # Iteration 1 hyperparams
  classifier: # Classifier training stage based on seen classes and discovered clusters
    network: 'multiclassnet' # Refer to networks.py for types of networks
    batch_size: 256 # Batch size for training
    num_epochs: 50 # Number of epochs for training
    learning_rate: 1.0e-4 # Learning rate for optimizer
    train_classes: 'Set1' # Set of classes to be used with their labels for training. Refer to class_list.yaml for set names
    transform: 'default' # Type of transform: Choose from 'default' for default transform or 'jpeg_blur' for augmented transformed images
    resize: 256 # Size of input image for network classification
  generate_features: # Generating features for all images
    extract_level: 'fc' # Feature extraction point from network, choose from: 'backbone' for output of backbone or 'fc' for output of fc layers
    eval_classes: 'Set1' # Set of classes whose images are to be discovered. Refer to class_list.yaml for set names. Labels used for evaluation of cluster metrics
  ood_detection: # Out of distribution detection of unlabeled images
    num_hashes: 2048 # Number of hashes to be used for Winner Take All (WTA) hash
    window_size: 2 # Window size for the hash
    normalize_type: 'mean_std' # Normalization type for features. Choose from 'mean_std' for mean-standard deviation, 'min_max' for minimum-maximum, 'none' for no normalization
    thresh_type: 'quantiles' # Type of thresholding for OOD based on array of scores: 'quantiles' for choosing {thresh_value}th quantile or 'mean_std' for choosing mean+{thresh_value}*std
    thresh_value: 0.9 # [0-1] for quantile type thresholding
  clustering: # K-means clustering
    num_clusters: 500 # Number of clusters for K-means
    max_iter: 300 # Number of iterations for clustering
    normalize_type: '_copy: ../../ood_detection/normalize_type' # Normalization type for features. Typically same as the type used for OOD detection.
    size_threshold: 0 # Minimum size of clusters (number of images) to be kept. Rest are discarded to undiscovered set
  merge: # Merge clusters based on a directed connected graph
    use_wta: False # Whether to use WTA hash for forming nearest neighbour graph
    connection_type: 'strong' # Type of connected graph, choose from: 'strong' for strongly connected graph and 'weak' for weakly connected graph for forming connected components
  refine:
    size_threshold: 100 # Minimum size of clusters (number of images) to be kept. Rest are discarded to undiscovered set
    purity_threshold: 0.5 # Clusters below this purity threshold (based on SVM predictions) are discarded entirely 
    kernel_type: 'rbf' # Type of SVM kernel, choose from 'rbf', 'linear'
  # addn_merge_refine_rounds: 0 # Number of additional rounds of merging and refining the discovery set

step2:
  _copy: '/step1' # Copy hyperparams from step 1
  classifier: 
    _copy: '/step1/classifier' 
    pretrained: True # Whether to start from previous iteration weights
    weighted: True # Whether to use weighted cross entropy loss due to imbalanced number of images in discovered clusters
    add_seen: False # Whether to include discovered images belonging to seen classes for training
    num_epochs: 100

step3:
  _copy: '/step2'

step4:
  _copy: '/step3'