DeepLearningExamples/TensorFlow2/Segmentation/Contrib/UNet3P/configs/config.yaml

# project root working directory, automatically read by hydra (.../UNet3P)
WORK_DIR: ${hydra:runtime.cwd}

DATA_PREPARATION:
  # unprocessed LiTS scan data paths, for custom data training skip this section details
  SCANS_TRAIN_DATA_PATH: "/data/Training Batch 2/"
  SCANS_VAL_DATA_PATH: "/data/Training Batch 1/"

  # Resize scans to model input size
  RESIZED_HEIGHT: ${INPUT.HEIGHT}
  RESIZED_WIDTH: ${INPUT.WIDTH}

  # Clip scans value in given range
  SCAN_MIN_VALUE: -200
  SCAN_MAX_VALUE: 250

DATASET:
  # paths should be relative from project root path
  TRAIN:
    IMAGES_PATH: "/data/train/images"
    MASK_PATH: "/data/train/mask"
  VAL:
    IMAGES_PATH: "/data/val/images"
    MASK_PATH: "/data/val/mask"


MODEL:
  # available variants are unet3plus, unet3plus_deepsup, unet3plus_deepsup_cgm
  TYPE: "unet3plus"
  WEIGHTS_FILE_NAME: model_${MODEL.TYPE}
  BACKBONE:
    # available variants are unet3plus, vgg16, vgg19
    TYPE: "vgg19"

DATA_GENERATOR_TYPE: "DALI_GENERATOR"  # options are TF_GENERATOR or DALI_GENERATOR
SEED: 5  # for result's reproducibility
VERBOSE: 1  # For logs printing details, available options are 0, 1, 2
DATALOADER_WORKERS: 3  # number of workers used for data loading
SHOW_CENTER_CHANNEL_IMAGE: True  # only true for UNet3+ for custom dataset it should be False

# Model input shape
INPUT:
  HEIGHT: 320
  WIDTH: 320
  CHANNELS: 3

# Model output classes
OUTPUT:
  CLASSES: 2


HYPER_PARAMETERS:
  EPOCHS: 100
  BATCH_SIZE: 16  # specify per gpu batch size
  LEARNING_RATE: 5e-5  # 0.1, 1e-3, 3e-4, 5e-5


CALLBACKS:
  # paths should be relative from project root path
  TENSORBOARD:
    PATH: "/checkpoint/tb_logs"

  EARLY_STOPPING:
    PATIENCE: 100

  MODEL_CHECKPOINT:
    PATH: "/checkpoint"
    SAVE_WEIGHTS_ONLY: True
    SAVE_BEST_ONLY: True

  CSV_LOGGER:
    PATH: "/checkpoint"
    APPEND_LOGS: False


PREPROCESS_DATA:
  RESIZE:
    VALUE: False  # if True, resize to input height and width
    HEIGHT: ${INPUT.HEIGHT}
    WIDTH: ${INPUT.WIDTH}

  IMAGE_PREPROCESSING_TYPE: "normalize"

  NORMALIZE_MASK:
    VALUE: False  # if True, divide mask by given value
    NORMALIZE_VALUE: 255

  SHUFFLE:
    TRAIN:
      VALUE: True
    VAL:
      VALUE: False


USE_MULTI_GPUS:
  VALUE: True  # If True use multiple gpus for training
  # GPU_IDS: Could be integer or list of integers.
  # In case Integer: if integer value is -1 then it uses all available gpus.
  # otherwise if positive number, then use given number of gpus.
  # In case list of Integers: each integer will be considered as gpu id
  # e.g. [4, 5, 7] means use gpu 5,6 and 8 for training/evaluation
  GPU_IDS: -1


OPTIMIZATION:
  AMP: True  # Automatic Mixed Precision(AMP)
  XLA: True  # Accelerated Linear Algebra(XLA)


# to stop hydra from storing logs files
# logs will be stored in outputs directory
defaults:
  - _self_
  - override hydra/hydra_logging: disabled
  - override hydra/job_logging: disabled

hydra:
  output_subdir: null