config.yaml

# Weights & Biases
USE_WANDB: True #@param {type:"boolean"}
WANDB_API_KEY: 'REPLACE_ME' #@param {type:"string"}

# Data
DATASET_PATH: 'REPLACE_ME' #@param {type:"string"}
MIN_VERTEBRAE_LEVEL: 8 #@param {type:"number"}
INPUT_DIM: 3 #@param [2, 3] {allow-input: true}
INPUT_SIZE: 64 #@param {type:"slider", min:32, max:112, step:4}

OVERSAMPLING: True
FOLD: 1

# Mask
# one of 'none', 'channel' (mask is additional input channel), 'apply' (single vertebra mask is applied to input),
# 'apply_all' (visible vertebra mask applied to input), 'crop' (mask is used to crop input)
MASK: 'none'

# Whether coordinates are provided in additional channels
# See Liu, Rosanne, et al. "An intriguing failing of convolutional neural networks and the CoordConv solution" (https://arxiv.org/pdf/1807.03247.pdf)
COORDINATES: False

# Whether to save the best-performing model (wrt validation F1)
SAVE_MODEL: False

# Data Augmentation
TRANSFORMS: ['modelsgenesis', 'intensity', 'spatial3d-simple']

# Task
TASK: "detection" # "detection", "grading", "simple_grading"
LOSS: 'binary_cross_entropy' # "ordinal_regression", "cross_entropy", "focal"

# Model
BATCH_SIZE: 16 #@param {type:"slider", min:32, max:512, step:32}
LEARNING_RATE: 0.001 #@param {type:"number"}
# AUTO_LR_FIND = True #@param {type: 'boolean'}
DROPOUT: 0.3 #@param {type:"number"}
WEIGHTED_LOSS: True #@param {type:"boolean"}
EARLY_STOPPING_PATIENCE: 40 #@param {type:"number"}

# Available backbones: 
#   - DenseNet121
#   - ModelsGenesis (3D)
#   - UNet3D
MODEL_NAME: 'UNet3D' #@param ["DenseNet121"] {allow-input: true}

# Identify all modules to freeze by name, e.g. ['down_tr64', 'down_tr128', 'down_tr256', 'down_tr512'] for Models Genesis/UNet3D
FROZEN_LAYERS: []