dataset:
  video_processor: VideoProcessor
  bert_name: bert-base-uncased
  meta_processor: MSRVTTMetaProcessor
  train_path: data/msrvtt/MSRVTT_train.csv
  dup: 20
  val_path: data/msrvtt/MSRVTT_JSFUSION_test.csv
  vfeat_dir: data/feat/feat_vtt_s3d
  text_processor: MSRVTTTextProcessor
  json_path: data/msrvtt/MSRVTT_data.json
  aligner: DSAligner
  num_iso_layer: 12
  max_video_len: 32
  max_len: 96
fairseq:
  common:
    tensorboard_logdir: run
    log_interval: 1000
    fp16: true
  dataset:
    num_workers: 4
    batch_size: 128
  optimization:
    lr:
    - 5.0e-05
    clip_norm: 2.0
    optimizer: adam
    adam_betas: (0.9, 0.98)
    lr_scheduler: polynomial_decay
    total_num_update: 1000000
    warmup_updates: 122
    weight_decay: 0.0
    ddp_backend: no_c10d
    max_epoch: 5
  checkpoint:
    restore_file: runs/mtm/vlm/checkpoint_best.pt
    reset_optimizer: true
    reset_dataloader: true
    reset_meters: true
    save_dir: runs/mtm/vlm/vttqa
task_type: sweep_small
model:
  model_cls: MMFusionJoint
  mm_encoder_cls: MMBertForJoint
  use_seg_emb: true
loss:
  loss_cls: V2TContraLoss