vrwkv_l_22kto1k_384.yaml

DATA:
  IMG_SIZE: 384
  BATCH_SIZE: 128 # * 8 gpus
  DATASET: 'imagenet'
  DATA_PATH: 'data/imagenet/'
  IMG_ON_MEMORY: False
MODEL:
  TYPE: 'vrwkv'
  PRETRAINED: './pretrained/vrwkv_l_in22k_192.pth'  # https://huggingface.co/OpenGVLab/Vision-RWKV/resolve/main/vrwkv_l_in22k_192.pth
  VRWKV:
    IMG_SIZE: 192
    PATCH_SIZE: 16
    IN_CHANNELS: 3
    OUT_INDICES: -1
    DROP_RATE: 0.
    EMBED_DIMS: 1024
    DEPTH: 24
    DROP_PATH_RATE: 0.1
    CHANNEL_GAMMA: 0.25
    SHIFT_PIXEL: 1
    SHIFT_MODE: 'q_shift'
    INIT_MODE: 'fancy'
    POST_NORM: True
    INIT_VALUES: 1e-5
    HIDDEN_RATE: 4
    FINAL_NORM: True
    INTERPOLATE_MODE: 'bicubic'
    HIDDEN_DIMS: 3072
TRAIN:
  EPOCHS: 20
  WARMUP_EPOCHS: 2
  WEIGHT_DECAY: 0.05
  BASE_LR: 5.0e-6
  WARMUP_LR: 5.0e-9
  MIN_LR: 5.0e-8
  CLIP_GRAD: 1.
  ACCUMULATION_STEPS: 1
  USE_CHECKPOINT: True
  EMA:
    ENABLE: True
    DECAY: 0.9999
  LR_SCHEDULER:
    NAME: 'cosine'
  LR_LAYER_DECAY: False
  LR_LAYER_DECAY_RATIO: 1.
AUG:
  COLOR_JITTER: 0.
  AUTO_AUGMENT: 'none'
  REPROB: 0.
  REMODE: 'pixel'
  RECOUNT: 1
  MIXUP: 0.8
  CUTMIX: 1.0
  CUTMIX_MINMAX: null
  MIXUP_PROB: 0.5
  MIXUP_SWITCH_PROB: 0.
  MIXUP_MODE: 'batch'
  RANDOM_RESIZED_CROP: True