point-lewm/config.json

{
  "_target_": "jepa.JEPA",
  "encoder": {
    "_target_": "pc_encoders.pointvit_encoder.PointViTEncoder",
    "embed_dim": 192,
    "in_channels": 3,
    "num_tokens": 256,
    "group_size": 32,
    "group_radius": 0.04,
    "group_max_neighbors": 128,
    "vit_size": "tiny",
    "vit_pretrained": false,
    "tokenizer_hidden": [
      128,
      256
    ],
    "use_feat": false,
    "norm_center": [
      0.98,
      0.0,
      0.17
    ],
    "norm_scale": 0.48
  },
  "predictor": {
    "_target_": "module.ARPredictor",
    "num_frames": 3,
    "input_dim": 192,
    "hidden_dim": 192,
    "output_dim": 192,
    "depth": 6,
    "heads": 16,
    "mlp_dim": 2048,
    "dim_head": 64,
    "dropout": 0.1,
    "emb_dropout": 0.0
  },
  "action_encoder": {
    "_target_": "module.Embedder",
    "input_dim": 10,
    "emb_dim": 192
  },
  "projector": {
    "_target_": "module.MLP",
    "input_dim": 192,
    "output_dim": 192,
    "hidden_dim": 2048,
    "norm_fn": {
      "_target_": "torch.nn.BatchNorm1d",
      "_partial_": true
    }
  },
  "pred_proj": {
    "_target_": "module.MLP",
    "input_dim": 192,
    "output_dim": 192,
    "hidden_dim": 2048,
    "norm_fn": {
      "_target_": "torch.nn.BatchNorm1d",
      "_partial_": true
    }
  }
}