training/configs/revcol_xlarge_1k_384_finetune.yaml

PRINT_FREQ: 30
SAVE_FREQ: 1
MODEL_EMA: True
DATA:
  IMG_SIZE: 384
  DATASET: imagenet
  PIPE_NAME: 'dpflow.silvia.imagenet.train.rand-re-jitt.384'
  NUM_WORKERS: 8
MODEL:
  TYPE: revcol_xlarge
  NAME: revcol_xlarge_1k_Finetune
  LABEL_SMOOTHING: 0.1
  NUM_CLASSES: 1000
REVCOL:
  INTER_SUPV: False
  SAVEMM: True
  FCOE: 3.0
  CCOE: 0.7
  DROP_PATH: 0.4
  HEAD_INIT_SCALE: 0.001
TRAIN:  
  EPOCHS: 30
  BASE_LR: 2e-5
  WARMUP_EPOCHS: 0
  WEIGHT_DECAY: 1e-8
  WARMUP_LR: 4e-6
  MIN_LR: 2e-7
  OPTIMIZER:
    NAME: 'adamw'
    LAYER_DECAY: 0.8

AUG:
  COLOR_JITTER: 0.0
# Use AutoAugment policy. "v0" or "original"
  AUTO_AUGMENT: 'rand-m9-mstd0.5-inc1'
# Random erase prob
  REPROB: 0.25
# Random erase mode
  REMODE: 'pixel'
# Random erase count
  RECOUNT: 1
# Mixup alpha, mixup enabled if > 0
  MIXUP:  0.0
# Cutmix alpha, cutmix enabled if > 0
  CUTMIX: 0.0
# Cutmix min/max ratio, overrides alpha and enables cutmix if set
  CUTMIX_MINMAX: None
# Probability of performing mixup or cutmix when either/both is enabled
  MIXUP_PROB: 0.0
# Probability of switching to cutmix when both mixup and cutmix enabled
  MIXUP_SWITCH_PROB: 0.0
# How to apply mixup/cutmix params. Per "batch", "pair", or "elem"
  MIXUP_MODE: 'batch'