DeePict_config.yaml

# General parameters for training and prediction:
dataset_table: "metadata.csv"   # Path to metadata csv file
output_dir: "memb/out"                     # Output directory
work_dir: "memb/work"                      # Destination directory of intermediate files
model_path: "memb_model.pth"               # Output model using .pth extension

# When running in the cluster:
cluster:
  logdir: "logs"

# Tomogram sets used for training or prediction
tomos_sets:
  training_list: ["TS_01", "TS_14", "TS_15"] # Tomograms in dataset_table for training ["tomo1", "tomo2", ...]
  prediction_list: ["TS_17"]             # Tomograms in dataset_table for prediction ["tomo1", "tomo2", ...]

cross_validation:
  active: false
  folds: 2
  statistics_file: "cv_statistics.csv"

training:
  active: True
  semantic_classes: ['actin']         # List of k classes that the network will learn, separated by commas
  processing_tomo: "filtered_tomo"            # Column name in dataset_table corresponding to raw tomo used for training
  box_size: 64                       # Box side of the partition
  min_label_fraction: 0.001          # Minimum label required in each box considered for the training partition
  overlap: 12                        # Thickness of overlap for training partition
  batch_size: 5                      # batch size for training
  force_retrain: false               # whether to rewrite model (set to false to not rewrite)

  # Unet architecture parameters
  unet_hyperparameters:
    depth: 4                         # unet depth (=number of maxpooling layers)
    initial_features: 32              # number of initial convolutions
    epochs: 1                      # training epochs
    train_split: 0.8                 # proportion of training (vs. validation) set, always between 0 and 1
    batch_norm: True                 # boolean value
    encoder_dropout: 0               # dropout for encoder path
    decoder_dropout: 0.2             # dropout for decoder path
    loss: "Dice"                     # Choose among: Dice (default), GeneralizedDice

  data_augmentation:
    rounds: 0                        # set to 0 to deactivate
    rot_angle: 180                   # rotation in degrees
    elastic_alpha: 0                 # elastic tranformation displacement parameter
    sigma_gauss: 1                   # variance associated to additive gaussian noise from 0 to sigma_gauss (usually < 5)
    salt_pepper_p: 0.01              # frequency of salt and pepper noise (uniformly distributed single pixel noise - between 0 and 1)
    salt_pepper_ampl: 0.1            # amplitude of salt and pepper noise (between 0 and 1)

prediction:
  active: true
  semantic_class: 'actin'          # Semantic class to be predicted

# Thresholding clustering and motl generation
postprocessing_clustering:
  active: True
  threshold: 0.5                         # Threshold for the probability score to make the predicted segmentation
  min_cluster_size: 100                  # Minimum number of voxels per cluster
  max_cluster_size: null                 # Maximum number of voxels per cluster
  clustering_connectivity: 3             # watershed algorithm clustering connectivity
  calculate_motl: False                  # Get the motl of centroids for each cluster
  ignore_border_thickness: [10, 20, 10]  # ignore border for motl generation if calculate_motl is True
  region_mask: 'no_mask'            # column name in metadata table for masking segmentation, e.g. lamella_file
  contact_mode: "intersection"           # "contact", "colocalization" or "intersection"
  contact_distance: 10

# For precision recall in particle picking
evaluation:
  particle_picking:
    active: false
    pr_tolerance_radius: 10              # radius in voxels for considering two coordinate corresp. to same particle
    statistics_file: "pr_statistics.csv" # statistics file where area under pr curve will be stored
  segmentation_evaluation:
    active: True
    statistics_file: "dice_eval.csv"     # statistics file where the dice coefficient will be stored

debug: True