vjepa2/configs/inference/vitg-384/ssv2.yaml at main · FJiangArthur/vjepa2 · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
cpus_per_task: 16
eval_name: video_classification_frozen
folder: /your_folder/evals/vitg-384/ssv2
mem_per_gpu: 220G
nodes: 16
num_workers: 8
resume_checkpoint: true
val_only: true
tag: ssv2-vitg16-384-64x2x3
tasks_per_node: 8
experiment:
  classifier:
    num_heads: 16
    num_probe_blocks: 4
  data:
    dataset_type: VideoDataset
    dataset_train: /your_data_path/ssv2_train_paths.csv
    dataset_val: /your_data_path/ssv2_val_paths.csv
    frame_step: 2
    frames_per_clip: 64
    num_classes: 174
    num_segments: 2
    num_views_per_segment: 3
    resolution: 384
  optimization:
    batch_size: 2
    multihead_kwargs:
    - final_lr: 0.0
      final_weight_decay: 0.0
      lr: 0.0
      start_lr: 0.0
      warmup: 0.0
      weight_decay: 0.0
    num_epochs: 20
    use_bfloat16: true
    use_pos_embed: false
model_kwargs:
  checkpoint: /your_vjepa2_checkpoints/vitg-384.pt
  module_name: evals.video_classification_frozen.modelcustom.vit_encoder_multiclip
  pretrain_kwargs:
    encoder:
      checkpoint_key: target_encoder
      img_temporal_dim_size: null
      model_name: vit_giant_xformers
      patch_size: 16
      tubelet_size: 2
      uniform_power: true
      use_rope: true
  wrapper_kwargs:
    max_frames: 128
    use_pos_embed: false