{ "model_type": "pinn_jepa_pretrain", "stage": "step1_pinn_pretrain", "fps": 30.0, "hidden_ratio": 2, "encoder": { "J": 17, "d_model": 256, "depth": 6, "d_state": 64, "temporal_kernel": 5, "mlp_ratio": 2, "dropout": 0.0, "use_checkpoint": false }, "input_layout": { "shape": "(B, T, J, 12)", "channels": ["p(3)", "v(3)", "a(3)", "j(3)"], "skeleton": "H36M17" }, "_comment": "These are architecture DEFAULTS. If the released checkpoint was trained with different values (d_model/depth/etc.), edit them here so load_state_dict shapes match. Set use_checkpoint=false for inference." }