2inf/rayzer_4

RayZer 4D — v12d and v12e

Training checkpoints and source-code snapshots for two 4D-ERayZer variants:

• v12d — erayzer_core.model.4d_erayzer_v12d.ERayZer4D trained on Dynamic-RE10K with the local-mean-loss variant (dre10k_freeze_dynrecon_localmeanloss_8gpu).
• v12e — erayzer_core.model.4d_erayzer_v12e.ERayZer4D — branch-color variant trained on Dynamic-RE10K (v12e_branch_color_dre10k_freeze_dynrecon_8gpu).

Repository layout

rayzer_4/
├── v12d/
│   ├── config.yaml                     # training config used for this run
│   ├── checkpoints/
│   │   ├── ckpt_0000000000005000.pt    # 5k, 10k, …, 30k steps (6 ckpts)
│   │   └── …
│   └── code/                           # source snapshot at train time
│       ├── erayzer_core/ data/ training_utils/ third_party/ …
│       ├── train.py
│       └── train_4d_v12d_dre10k_freeze_dynrecon_localmeanloss.yaml
└── v12e/
    ├── config.yaml
    ├── checkpoints/                    # 5k, 10k, …, 40k (8 ckpts)
    └── code/                           # same layout as v12d/code/

Each checkpoint is ~3.2 GB. wandb logs, evaluation dumps, __pycache__, and vendored .git histories have been stripped from the code snapshots.

Loading

import torch, yaml
from omegaconf import OmegaConf

cfg = OmegaConf.load("v12d/config.yaml")

# Make the snapshotted source tree importable:
import sys; sys.path.insert(0, "v12d/code")

from erayzer_core.model import __dict__ as models  # noqa
# v12d → models["4d_erayzer_v12d"] exposes the ERayZer4D class
# v12e → models["4d_erayzer_v12e"] exposes the ERayZer4D class
ModelMod = __import__("erayzer_core.model.4d_erayzer_v12d", fromlist=["ERayZer4D"])
model = ModelMod.ERayZer4D(cfg)

sd = torch.load("v12d/checkpoints/ckpt_0000000000030000.pt",
                map_location="cpu", weights_only=False)
model.load_state_dict(sd["model"], strict=False)
model.eval()

The config.yaml under each version dir matches what train.py consumed during that run. Inference configs for downstream gradio / viewer apps live under code/config/.

Related

• Upstream code base: this repo is a subset of the 4D_RAVI working tree at the commits these runs were launched from.
• Raw training data mirrors: 2inf/kitti-raw, 2inf/nuscenes-raw, 2inf/waymo-raw (the 4D-RayZer models here are trained on Dynamic-RE10K, not on those three).

License

Research use only. See the upstream WildRayZer / RayZer licenses for details.