ProWorld / train_hyperbolic_scene.py
lzhts1's picture
Upload 3 files
26eef82 verified
Raw
History Blame Contribute Delete
919 Bytes
"""
Train the HyperbolicJEPA variant on OGBench Scene.
Example:
MUJOCO_GL=egl python train_hyperbolic_scene.py \
trainer.accelerator=gpu \
trainer.devices=1 \
trainer.max_epochs=10 \
loader.batch_size=128 \
subdir=ogbench_scene/hyperbolic_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_hyperbolic import run_hyperbolic_training
@hydra.main(
version_base=None,
config_path="./config/train",
config_name="lewm_hyperbolic_scene",
)
def run(cfg):
run_hyperbolic_training(cfg)
@record
def _main():
faulthandler.enable(all_threads=True)
run()
if __name__ == "__main__":
_main()