# CartoCell inference workflow compatible with BiaPy version 3.6.8

SYSTEM:
    NUM_CPUS: -1

# Problem definition
PROBLEM:
    TYPE: INSTANCE_SEG          # instace segmentation workflow
    NDIM: 3D                    # images are 3D
    INSTANCE_SEG:
        DATA_CHANNELS: BCM              # problem representation: binary masks + contours + foreground mask
        DATA_REMOVE_BEFORE_MW: True     # remove objects before watershed
        DATA_REMOVE_SMALL_OBJ_BEFORE: 5 # size of small objects to be removed before watershed

# Data paths and management
DATA: 
    PATCH_SIZE: (80, 80, 80, 1) # model input size (z, x, y, channels)    
    TRAIN:                                                                                                              
        PATH: /path/to/data     # full path to folder with training raw images
        GT_PATH: /path/to/data  # full path to folder with training label images
        IN_MEMORY: True         # load tranining data into RAM (faster) or not (save memory)
    VAL:
        FROM_TRAIN: False       # do not use random samples from training as validation set
        PATH: /path/to/data     # full path to folder with validation raw images
        GT_PATH: /path/to/data  # full path to folder with validation label images
    TEST:                                                                                                               
        PATH: /path/to/data     # full path to folder with test raw images
        GT_PATH: /path/to/data  # full path to folder with test label images
        IN_MEMORY: True         # load teset data into RAM (faster) or not (save memory)
        LOAD_GT: True           # check if test labels available
        PADDING: (16,16,16)     # padding to be used during test (z, x, y)
    REFLECT_TO_COMPLETE_SHAPE: True # use mirror padding when needed

# Data augmentation
AUGMENTOR:
    ENABLE: True                    # wether to use data augmentation or not
    RANDOM_ROT: True                # random rotations
    VFLIP: True                     # vertical flips
    HFLIP: True                     # horizontal flips
    ZFLIP: True                     # z-axis flips
    ELASTIC: True                   # elastic deformations
    BRIGHTNESS: True                # use brightness augmentation
    BRIGHTNESS_FACTOR: (0.1, 0.15)  # minimum and maximum brightness factors to apply

# Deep learning model configuration
MODEL:
    ARCHITECTURE: resunet           # architecture (ResU-Net)
    FEATURE_MAPS: [52, 68, 84]      # number of feature maps at each level
    Z_DOWN: [1,1]                   # do not downscale in z-axis after each level
    DROPOUT_VALUES: [0.1, 0.1, 0.1] # dropout values at each level
    LOAD_CHECKPOINT: True           # load pretrained model (path to the model is needed)

# Path to the pretrained model
PATHS:
    CHECKPOINT_FILE: /home/user/cartocell_model_latest.pth

# Model training
TRAIN:
    ENABLE: False       # do not perform training (only inference)
    OPTIMIZER: ADAMW    # optimizer
    LR: 1.E-3           # learning rate (maximum value for one-cycle scheduler)
    BATCH_SIZE: 2       # batch size
    EPOCHS: 30          # number of epochs
    PATIENCE: 30        # patience
    LR_SCHEDULER:
        NAME: 'onecycle'    # use learning one-cycle rate scheduler

# Loss function
LOSS:
    CLASS_REBALANCE: True # give the same weight to all channels (in BCM)

# Model testing
TEST:
    ENABLE: True        # perform testing (inference)
    AUGMENTATION: False # do not perform test-time augmentation
    FULL_IMG: False     # do not process full images (use patches)
    POST_PROCESSING:
        VORONOI_ON_MASK: True   # apply Voronoi algorithm to improve masks
        CLEAR_BORDER: False     # do not remove objects in the border