{
 "model": "vjepa2_1-vitb-384",
 "source": "/home/overseer2/workdir/jepa.cpp/models/vjepa2_1/vjepa2_1_vitb_dist_vitG_384.pt",
 "dtype": "float32",
 "npy_dtype_note": "all tensors float32 C-order except frames_u8 (uint8) and top5_idx (int64)",
 "framework": {
  "torch": "2.13.0+cpu",
  "transformers": "5.16.1",
  "python": "3.12.13",
  "threads": 32,
  "cpu": "x86_64"
 },
 "created": "2026-08-30T22:33:17+0300",
 "source_url": "https://dl.fbaipublicfiles.com/vjepa2/vjepa2_1_vitb_dist_vitG_384.pt",
 "code": "https://github.com/facebookresearch/vjepa2 (hubconf vjepa2_1_vit_base_384, app/vjepa_2_1/models/vision_transformer.py), local clone at /home/overseer2/workdir/jepa.cpp/tmp/vjepa2-src",
 "checkpoint": {
  "top_level_keys": [
   "encoder",
   "predictor",
   "opt",
   "scaler",
   "ema_encoder",
   "epoch",
   "loss",
   "batch_size",
   "world_size",
   "lr"
  ],
  "encoder_key_used": "ema_encoder",
  "key_cleanup": "strip 'module.' and 'backbone.'",
  "epoch": 40
 },
 "hparams": {
  "embed_dim": 768,
  "n_layer": 12,
  "n_head": 12,
  "ffn_dim": 3072,
  "patch_size": 16,
  "tubelet_size": 2,
  "img_size": 384,
  "n_frames": 64,
  "ln_eps": 1e-06,
  "act": "gelu_erf",
  "qkv_bias": true,
  "cls_token": false,
  "pos_type": "rope3d",
  "rope_theta": 10000.0,
  "rope_interpolate": true,
  "rope_pretrained_grid": 16,
  "rope_dims_per_axis": [
   20,
   20,
   20
  ],
  "modality_embed": true,
  "image_patch_embed": true,
  "hier_layers": [
   2,
   5,
   8,
   11
  ],
  "out_layers_distillation": [
   2,
   5,
   8,
   11
  ],
  "pred": {
   "embed_dim": 384,
   "n_layer": 12,
   "n_mask_tokens": 8,
   "teacher_embed_dim": 1664
  }
 },
 "preprocessing": {
  "description": "VJEPA2VideoProcessor: uint8 RGB frames (T,H,W,3) -> per-frame resize so the SHORT side = 438 (= int(crop*256/224)), aspect kept, torchvision.transforms.v2.functional.resize on the uint8 CHW tensor with BILINEAR + antialias=True (result rounded back to uint8) -> center crop 384x384 (top = int((H-384)/2), left = int((W-384)/2)) -> x/255 -> ImageNet mean/std -> pixel_values_videos [B, T, 3, H, W]",
  "resize": {
   "shortest_edge": 438,
   "resample": "bilinear",
   "antialias": true,
   "on_dtype": "uint8"
  },
  "center_crop": 384,
  "rescale": 0.00392156862745098,
  "mean": [
   0.485,
   0.456,
   0.406
  ],
  "std": [
   0.229,
   0.224,
   0.225
  ],
  "frame_sampling": "idx = round(linspace(0, T_total-1, n)) over all decoded frames (PyAV rgb24); indices stored per sample",
  "resampling_note": "resampling = torchvision.transforms.v2.functional.resize on the uint8 CHW tensor with antialias=True (result rounded back to uint8), exactly what transformers>=5 image/video processors (TorchvisionBackend) do; verified to 2.4e-7 against a re-implementation, whereas PIL resampling differs by up to 1.8e-2 (1-2 uint8 levels)",
  "processor": {
   "class": "VJEPA2VideoProcessor",
   "backend": "torchvision",
   "config": {
    "crop_size": {
     "height": 384,
     "width": 384
    },
    "size": {
     "shortest_edge": 438
    },
    "do_resize": true,
    "resample": 2,
    "do_rescale": true,
    "rescale_factor": 0.00392156862745098,
    "do_normalize": true,
    "do_center_crop": true,
    "return_metadata": false,
    "image_mean": [
     0.485,
     0.456,
     0.406
    ],
    "image_std": [
     0.229,
     0.224,
     0.225
    ],
    "video_processor_type": "VJEPA2VideoProcessor"
   }
  },
  "note": "No HF processor exists for 2.1; we reuse VJEPA2VideoProcessor(crop_size=384) (short side 438, bilinear+antialias on uint8, center crop 384, ImageNet stats) which mirrors Meta's eval transform (Resize(int(size*256/224)) + CenterCrop(size)).  Images go through the same processor as a 1-frame video.  The tensor is then permuted to NCTHW because the Meta encoder takes (B, C, T, H, W)."
 },
 "outputs": {
  "input": "NCTHW float32; video: T=16 -> tubelets of 2 -> 8 x 24 x 24 = 4608 tokens; image: T=1 -> patch_embed_img (1x16x16) + img_mod_embed -> 24 x 24 = 576 tokens",
  "last_hidden_state": "encoder(x)[0] in default inference mode = norms_block[-1](x after the last block) [N, 768]; token order t-major then h then w",
  "pooled_mean": "mean over tokens"
 },
 "timing_s": {
  "model_load": 1.479,
  "forward_total": 2.035,
  "forward_mean": 0.509,
  "wall_total": 2.644
 },
 "samples": [
  {
   "name": "archery_f16",
   "media": "archery.mp4",
   "timing_s": {
    "decode_s": 0.331,
    "preprocess_s": 0.0162,
    "forward_s": 0.8974
   },
   "tensors": {
    "frames_u8": {
     "file": "archery_f16.frames_u8.npy",
     "shape": [
      16,
      360,
      480,
      3
     ],
     "dtype": "uint8",
     "layout": "THWC uint8"
    },
    "input": {
     "file": "archery_f16.input.npy",
     "shape": [
      1,
      3,
      16,
      384,
      384
     ],
     "dtype": "float32",
     "layout": "NCTHW"
    },
    "last_hidden_state": {
     "file": "archery_f16.last_hidden_state.npy",
     "shape": [
      4608,
      768
     ],
     "dtype": "float32",
     "layout": "[N_tokens, D] t-major,h,w"
    },
    "pooled_mean": {
     "file": "archery_f16.pooled_mean.npy",
     "shape": [
      768
     ],
     "dtype": "float32",
     "layout": "[D]"
    }
   },
   "frames": 16,
   "frame_indices": [
    0,
    20,
    40,
    60,
    80,
    100,
    120,
    140,
    159,
    179,
    199,
    219,
    239,
    259,
    279,
    299
   ],
   "n_frames_total": 300,
   "fps": 29.97002997002997,
   "frame_size_hw": [
    360,
    480
   ],
   "path": "video"
  },
  {
   "name": "bowling_f16",
   "media": "bowling.mp4",
   "timing_s": {
    "decode_s": 0.1574,
    "preprocess_s": 0.0388,
    "forward_s": 0.9183
   },
   "tensors": {
    "frames_u8": {
     "file": "bowling_f16.frames_u8.npy",
     "shape": [
      16,
      270,
      480,
      3
     ],
     "dtype": "uint8",
     "layout": "THWC uint8"
    },
    "input": {
     "file": "bowling_f16.input.npy",
     "shape": [
      1,
      3,
      16,
      384,
      384
     ],
     "dtype": "float32",
     "layout": "NCTHW"
    },
    "last_hidden_state": {
     "file": "bowling_f16.last_hidden_state.npy",
     "shape": [
      4608,
      768
     ],
     "dtype": "float32",
     "layout": "[N_tokens, D] t-major,h,w"
    },
    "pooled_mean": {
     "file": "bowling_f16.pooled_mean.npy",
     "shape": [
      768
     ],
     "dtype": "float32",
     "layout": "[D]"
    }
   },
   "frames": 16,
   "frame_indices": [
    0,
    10,
    20,
    30,
    40,
    50,
    60,
    70,
    79,
    89,
    99,
    109,
    119,
    129,
    139,
    149
   ],
   "n_frames_total": 150,
   "fps": 15.0,
   "frame_size_hw": [
    270,
    480
   ],
   "path": "video"
  },
  {
   "name": "coco_000000000139",
   "media": "coco_000000000139.jpg",
   "timing_s": {
    "preprocess_s": 0.0025,
    "forward_s": 0.1112
   },
   "tensors": {
    "input": {
     "file": "coco_000000000139.input.npy",
     "shape": [
      1,
      3,
      1,
      384,
      384
     ],
     "dtype": "float32",
     "layout": "NCTHW (T=1, image path)"
    },
    "last_hidden_state": {
     "file": "coco_000000000139.last_hidden_state.npy",
     "shape": [
      576,
      768
     ],
     "dtype": "float32",
     "layout": "[N_tokens, D] h-major"
    },
    "pooled_mean": {
     "file": "coco_000000000139.pooled_mean.npy",
     "shape": [
      768
     ],
     "dtype": "float32",
     "layout": "[D]"
    }
   },
   "image_size_hw": [
    426,
    640
   ],
   "path": "image"
  },
  {
   "name": "coco_000000000285",
   "media": "coco_000000000285.jpg",
   "timing_s": {
    "preprocess_s": 0.0016,
    "forward_s": 0.1086
   },
   "tensors": {
    "input": {
     "file": "coco_000000000285.input.npy",
     "shape": [
      1,
      3,
      1,
      384,
      384
     ],
     "dtype": "float32",
     "layout": "NCTHW (T=1, image path)"
    },
    "last_hidden_state": {
     "file": "coco_000000000285.last_hidden_state.npy",
     "shape": [
      576,
      768
     ],
     "dtype": "float32",
     "layout": "[N_tokens, D] h-major"
    },
    "pooled_mean": {
     "file": "coco_000000000285.pooled_mean.npy",
     "shape": [
      768
     ],
     "dtype": "float32",
     "layout": "[D]"
    }
   },
   "image_size_hw": [
    640,
    586
   ],
   "path": "image"
  }
 ]
}