{
  "architectures": [
    "STP"
  ],
  "model_type": "stp",
  "image_size": 224,
  "patch_size": 16,
  "encoder": {
    "embed_dim": 768,
    "depth": 12,
    "num_heads": 12,
    "pos_embed": "fixed sincos"
  },
  "spatial_decoder": {
    "embed_dim": 512,
    "depth": 8,
    "num_heads": 16
  },
  "temporal_decoder": {
    "embed_dim": 512,
    "depth": 8,
    "num_heads": 16,
    "type": "self-attn + cross-attn(context=Z_c) + FFN"
  },
  "mask_ratio_current": 0.75,
  "mask_ratio_future": 0.95,
  "frame_interval": 16,
  "num_parameters": 146336000,
  "source_epoch": 49,
  "pretraining": {
    "optimizer": "AdamW",
    "blr": 0.00015,
    "weight_decay": 0.05,
    "betas": [
      0.9,
      0.95
    ],
    "epochs": 50,
    "warmup_epochs": 5,
    "effective_batch_size": 4096,
    "augmentation": "RandomResizedCrop(0.8, 1.0)",
    "dataset": "EgoClip (Ego4D)",
    "normalize_mean": [
      0.485,
      0.456,
      0.406
    ],
    "normalize_std": [
      0.229,
      0.224,
      0.225
    ]
  },
  "usage": "freeze the encoder, take the [CLS] token via forward_features()",
  "paper": "International Journal of Computer Vision (2026), doi:10.1007/s11263-026-02948-3"
}