Skip to content

README_OCCSTRESS.md

Latest commit

 

History

History
46 lines (39 loc) · 2.52 KB

README_OCCSTRESS.md

File metadata and controls

46 lines (39 loc) · 2.52 KB
Raw

I2-World OccStress Adapter

This snapshot contains the two-stage nuScenes adapter and the Waymo/CARLA scene-shard wrapper. A fresh environment passed small A100 tests on all three datasets. Metrics match the pre-migration snapshot; Waymo/CARLA confusion matrices also match. This is not full-paper reproduction. See release status.

The existing interface exports tokens before forecasting. In an isolated I2-World environment, set the shared roots at the release root, then:

cd "$OCCSTRESS_CODE_ROOT/EXIST/4D/II-World"
export PROTOCOL_NAME=clean_H4_F6_val_backbone
export PROTOCOL_PATH="$OCCSTRESS_DATA_ROOT/protocols/manual/OccStress-nuScenes/clean/H4_F6_val_backbone.pkl"
export IIWORLD_BASE_INFO=/path/to/world-nuscenes_infos_trainval.pkl
export OCCSTRESS_IIWORLD_SAVE_ROOT=/path/to/temporary/protocol-tokens
export OCCSTRESS_IIWORLD_TOKEN_ROOT="$OCCSTRESS_IIWORLD_SAVE_ROOT/token_4f"
export IIWORLD_CLEAN_TOKEN_ROOT=/path/to/clean-token_4f
python tools/test.py configs/scene_tokenizer/ii_scene_tokenizer_occstress.py \
  /path/to/tokenizer.pth --eval mIoU
python tools/test.py configs/world_model/ii_generate_world_occstress.py \
  /path/to/world_model.pth --eval mIoU

Clean tokens must be prepared with the matching tokenizer/config. Token caches are not distributed as checkpoint or source code assets. The two configs now consume the environment variables above, and their dataset import names match the packaged module.

For Waymo/CARLA use tools/test_waymo_occstress.py through the formal selector in docs/FORMAL_EVALUATION.md. It exports current tokens per protocol/scene shard, uses a shared future-token cache, streams confusion matrices instead of retaining predictions, and removes temporary current tokens after success. Traffic future tokens are isolated from unmirrored targets; config, adapter, base and checkpoint hashes participate in the future cache identity.

The explicit --dataset carla must be paired with both CARLA configs. The formal selector supplies them. --keep-token-cache is debugging-only. Use a distinct work root per concurrent shard, and merge confusion counts before computing scores. The correction to zero-IoU class handling is described in the formal guide; do not silently mix historical native log scores with present-class summaries.

The MIT notice is restored from experimental base 2b487f8; later upstream HEAD deleted it. See the license audit for the version-specific evidence, not a blanket grant for future updates.