kinetics600_ft.yaml

DATA:
    ROOT: 'data/'
    TRAIN_FILE: 'data/k600_train.txt'
    VAL_FILE: 'data/k600_val.txt'
    DATASET: kinetics600
    NUM_FRAMES: 16
    TUBELET_SIZE: 1
    NUM_CLASSES: 600
    LABEL_LIST: 'labels/kinetics600_labels.csv'
MODEL:
    ARCH: eva-1b/14
    DROP_PATH_RATE: 0.2
    PRETRAINED: /path/to/eva_video_k722.pth  # download from https://huggingface.co/BAAI/EVA/blob/main/eva_video_k722.pth
    PRETRAINED_MODE: 'eva_k722'
    CLASS_MAPPING: 'data/k722_to_k600_mapping.npy'
AUG:
    LABEL_SMOOTH: 0.1
    MIXUP: 0.
    CUTMIX: 0.
    COLOR_JITTER: 0.8
    GRAY_SCALE: 0.2
TRAIN:
    BATCH_SIZE: 1
    ACCUMULATION_STEPS: 4
    BETAS: (0.9, 0.98)
    EPS: 1e-6
    LR: 1.e-6
    WARMUP_START_LR: 1.e-6
    LR_MIN: 1.e-6
    LAYER_WISE_DECAY: 0.95
    WEIGHT_DECAY: 0.05
    EPOCHS: 2
    WARMUP_EPOCHS: 0.
    USE_CHECKPOINT: True
MYCLIP:
    USE_LEARNABLE_POS_EMB: True
    STOP_GRAD_CONV1: False
    USE_MEAN_POOLING: True
    USE_TEXT_EMBED: False
    FT_FRAMES: 8
    FT_IMAGE_SIZE: 224