train_scene.py

"""
Train the original LeWM model on OGBench Scene.

Example:
MUJOCO_GL=egl python train_scene.py \
  trainer.accelerator=gpu \
  trainer.devices=1 \
  trainer.max_epochs=10 \
  loader.batch_size=128 \
  subdir=ogbench_scene/lewm_scene \
  wandb.enabled=False

If your local cache uses a different dataset layout, override the dataset name:
  data.dataset.name=/path/to/visual-scene-play-v0.h5
"""

import faulthandler

import hydra

try:
    from torch.distributed.elastic.multiprocessing.errors import record
except Exception:
    def record(fn):
        return fn

from train import run_training


@hydra.main(
    version_base=None,
    config_path="./config/train",
    config_name="lewm_scene",
)
def run(cfg):
    run_training(cfg)


@record
def _main():
    faulthandler.enable(all_threads=True)
    run()


if __name__ == "__main__":
    _main()