project_dir: mfmmlm
run_task:
  - how2.yaml
  - [vtt.yaml, vttcap.yaml, vttqa.yaml, youcook.yaml, youcookcap.yaml, crosstask.yaml, coin.yaml]
base_dir: task
task_group:
  pretrain:
    task_list:
      - how2.yaml
    dataset:
      subsampling: 32
      sampled_min_len: 10
      sampled_max_len: 64
      max_video_len: 32
      max_len: 96
      aligner: MFMMLMAligner
      lazy_vfeat_mask: True
      mfm_probability: 0.15
      mlm_probability: 0.15
      mm_prob: 0.5
    model:
      model_cls: MMFusionMFMMLM
      mm_encoder_cls: MMFusionForMFMMLM
    loss:
      loss_cls: MFMMLM
    fairseq:
      common:
        fp16: true
      dataset:
        batch_size: 256
      optimization:
        max_epoch: 15     
  finetune:
    task_list:
      - vtt.yaml
      - vttqa.yaml
      - youcook.yaml
      - youcookcap.yaml
      - crosstask.yaml
      - coin.yaml
    dataset:
      max_video_len: 32
      max_len: 96
    fairseq:
      common:
        fp16: true
    # do not write any model or loss here (they are expected to be fixed in mmfusion).
  test:
    task_list:
      - test_vtt.yaml
      - test_vttqa.yaml
      - test_youcook.yaml
      - test_youcookcap.yaml
      - test_crosstask.yaml
      - test_crosstask_zs.yaml
      - test_coin.yaml
    dataset:
      max_video_len: 32
      max_len: 96