model:
    pretrain: ckpt/charades_ego.pt
    freeze_vis_backbone: true
    freeze_txt_backbone: true
    num_frames: 16
    text_prompt:
        n_ctx: 8
        use_bank: true
    visual_prompt:
        num_layers: 12
        prompt_dim: 512
        num_tokens: 128
        deep: true
        deep_shared: false
        split_st: false
        pt_spt: true
        pt_tmp: false
        style: VoP_c_pool  # VoP_c: prompts are generated by context fusion; frame-specific attention
        n_seg: 16     # number of segments per video (n_seg=clip_length -> 1 frame/seg)
        K_s: 8        # boundary of intra-frame/inter-frame attention (VoP_f+c)
        pool:
            size: 10

data:
    dataset: charades_ego
    #root: /data/CharadesEgo/CharadesEgo_v1_480
    #metadata_val: /data/CharadesEgo/CharadesEgo/CharadesEgo_v1_test_only1st.csv        # all testing data
    root: data/charades_ego/video
    metadata_val: data/charades_ego/csv/{}.csv
    label_map: meta/charades_ego/charades_ego.json
    clip_length: 16
    sparse_sample: true