config.yml

# ----------------
# ---- Common ----
# ----------------
SOURCE:
  'Unity'
TARGET:
  'Real'
NUM_WORKERS:
  16
NUM_CLASSES:
  2 # it was 19
EXP_NAME:
  ''
EXP_ROOT:
  project_root / 'experiments'
EXP_ROOT_SNAPSHOT:
  '' # osp.join(cfg.EXP_ROOT, 'snapshots')
EXP_ROOT_LOGS:
  '' # osp.join(cfg.EXP_ROOT, 'logs')
GPU_ID:
  0

# ----------------
# ----- Train-----
# ----------------
TRAIN:
  INPUT_SIZE_SOURCE: !!python/tuple [256,256]
  INPUT_SIZE_TARGET: !!python/tuple [256,256]
  MODEL:
    'DeepLabv2'
  MULTI_LEVEL:
    True
  RESTORE_FROM:
    ''
  LEARNING_RATE:
    2.5e-4
  MOMENTUM:
    0.9
  WEIGHT_DECAY:
    0.0005
  POWER:
    0.9
  LAMBDA_SEG_MAIN:
    1.0
  LAMBDA_SEG_AUX:
    0.1
  DA_METHOD:
    'AdvEnt'
  # Adversarial training params
  LEARNING_RATE_D:
    0.0001
  LAMBDA_ADV_MAIN:
    0.001
  LAMBDA_ADV_AUX:
    0.0002
  # MinEnt params
  LAMBDA_ENT_MAIN:
    0.001
  LAMBDA_ENT_AUX:
    0.0002
  # Other params
  MAX_ITERS:
    250000
  EARLY_STOP:
    30000
  SAVE_PRED_EVERY:
    2000
  SNAPSHOT_DIR:
    ''
  RANDOM_SEED:
    1234
  tests_per_epoch:
    4
  SAVE_IMAGE_PRED:
    100

# ----------------
# ----- Test -----
# ----------------
TEST:
  MODE:
    'best' # {'single', 'best'}
  Model:
    TEST.MODEL:
      ('DeepLabv2',)
    MODEL_WEIGHT:
      (1.0,)
    MULTI_LEVEL:
      (True,)
    RESTORE_FROM:
      ''
    SNAPSHOT_DIR:
      '' # used in 'best' mode
    SNAPSHOT_STEP:
      2000  # used in 'best' mode
    SNAPSHOT_MAXITER:
      120000  # used in 'best' mode
    # Test sets
    SET_TARGET:
      'val'
    INPUT_SIZE_TARGET: !!python/tuple [256,256]
    OUTPUT_SIZE_TARGET: !!python/tuple [256,256]
    WAIT_MODEL:
      True
    overlay:
      True
    store_images:
      False
    fixed_test_size:
      True
    test_iter:
      30000


# ----------------
# --- Comet_ML ---
# ----------------
workspace:
  'tianyu-z'
project_name:
  'advent-ccai'

# ----------------
# ----- Data -----
# ----------------
data:
  use_real: true
  files: # if one is not none it will override the dirs location
    base: "/network/tmp1/ccai/data/mask_generation/11K" # ! Check output_dir
    train: train.json
    val: test.json

  real_files:
    base: "/network/tmp1/ccai/data/mask_generation/real"
    train: train.json
    val: test.json

  img_size: 256

  loaders:
    batch_size: 40 # 40, max=93
    shuffle: True
    num_workers: 4
  transforms:
    - name: hflip
      ignore: true
      p: 0.5
    - name: resize
      ignore: false
      new_size: 256 #! Make sure this matches opts.data.img_size
    - name: crop
      ignore: false
      height: 224
      width: 224
    - name: resize # ? this or change generator's output? Or resize larger then crop to 256?
      ignore: false
      new_size: 256 #! Make sure this matches opts.data.img_size