16_16_text_tuned_clip.yaml

DATA:
    ROOT: '/PATH/TO/videos/folder'
    TRAIN_FILE: 'datasets_splits/k400_splits/train.txt'
    VAL_FILE: 'datasets_splits/k400_splits/val.txt'
    DATASET: kinetics400
    NUM_FRAMES: 16
    NUM_CLASSES: 400
    LABEL_LIST: 'labels/kinetics_400_labels.csv'
MODEL:
    ARCH: ViT-B/16
TRAIN:
    BATCH_SIZE: 4 # BS 256
    ACCUMULATION_STEPS: 8
    LR: 8e-06
TEST:
    MULTI_VIEW_INFERENCE: True
TRAINER:
  ViFi_CLIP:
    ZS_EVAL: False # Make True only during test mode to evaluate zero-shot vanilla CLIP performance
    USE: "text" # text refers to tuning only text encoder of CLIP