1X World Model Challenge

Code for 1X world model challenge, currently using a transformer for compression.

Config for training:

cd /home/ubuntu/world-model/1x-world-model && PYTHONPATH="/home/ubuntu/world-model/1x-world-model/src" python3 src/world_model/__main__.py fit \
  --data.num_workers=16 \
  --data.batch_size=8 \
  --trainer.accelerator=gpu \
  --trainer.devices=1 \
  --trainer.precision=16-mixed \
  --trainer.max_epochs=10 \
  --trainer.logger=WandbLogger \
  --trainer.logger.project=my-world-model \
  --trainer.logger.name=baseline-full-dataset

Evaluate CE and Export Submission

Prereqs: ensure your Python path includes src and you have a Lightning checkpoint.

1. Quick CE evaluation on val (adjust --max-samples as needed):

PYTHONPATH=src python3 -m world_model.eval_and_export my-world-model/2f4f2142/checkpoints/epoch=3-step=628300.ckpt data --eval --max-samples 50 --k 500 --device cuda

2. Export submission NPZs (indices/values of shape (3,32,32,500)) and README.txt:

PYTHONPATH=src python3 -m world_model.eval_and_export my-world-model/2f4f2142/checkpoints/epoch=3-step=628300.ckpt data --export-dir out_submission --max-samples 450 --k 500 --device cuda \
  --team "your-team" --authors "you" --email "[email protected]" --institution "org" --country "XX"

3. Create a flat submission.zip and validate with the official validator:

(cd out_submission && zip -q -r ../submission.zip .)
python3 data/test_v2.0/validate_submission.py submission.zip --num-samples 450

If you see warnings about sub-directories, make sure you zipped from within out_submission so the zip is flat.