vjepa2/configs/inference/vitg-384/diving48.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
nodes: 8
tasks_per_node: 8
cpus_per_task: 12
mem_per_gpu: 200G
tag: diving48-vitg16-384-32x4x3
eval_name: video_classification_frozen
folder: /your_folder/evals/vitg-384/diving48
resume_checkpoint: true
val_only: true
experiment:
  classifier:
    num_probe_blocks: 4
    num_heads: 16
  data:
    dataset_type: VideoDataset
    dataset_train: /your_data_dir/diving48/annotations/Diving48_train_paths.csv
    dataset_val: /your_data_dir/diving48/annotations/Diving48_test_paths.csv
    num_classes: 48
    resolution: 384
    frames_per_clip: 32
    frame_step: 2
    num_segments: 4
    num_views_per_segment: 3
  optimization:
    use_pos_embed: false
    num_epochs: 100
    batch_size: 2
    use_bfloat16: true
    multihead_kwargs:
      - weight_decay: 0.0
        final_weight_decay: 0.0
        lr: 0.0
        start_lr: 0.0
        final_lr: 0.0
        warmup: 0.0
model_kwargs:
  checkpoint: /your_vjepa2_checkpoints/vitg-384.pt
  module_name: evals.video_classification_frozen.modelcustom.vit_encoder_multiclip_multilevel
  wrapper_kwargs:
    max_frames: 128
    use_pos_embed: false
    out_layers: [24, 29, 34, 39]
  pretrain_kwargs:
    encoder:
      model_name: vit_giant_xformers
      checkpoint_key: target_encoder
      tubelet_size: 2
      patch_size: 16
      uniform_power: true
      use_rope: true