1. Download nuScenes V1.0 full dataset data HERE. Folder structure:
nuscenes/
|── maps/
│ ├── expansion/
├── samples/
├── sweeps/
├── lidarseg/
├── v1.0-test/
├── v1.0-trainval/
2. Set up the environment, edit the ENV.py:
os.environ['TORCH_HOME'] = $PATH_TO_TORCH_HOME
os.environ['TRANSFORMERS_CACHE'] = $PATH_TO_TRANSFORMERS_CACHE
os.environ['HUGGINGFACE_HUB_CACHE'] = $PATH_TO_HUGGINGFACE_HUB_CACHE
os.environ['XDG_CACHE_HOME'] = $PATH_TO_XDG_CACHE
3. Convert raw nuScenes to our format for fast training (v1.0-trainval take days, v1.0-mini take hours):
python ./dreamer-datasets/dd_scripts/converters/nuscenes_converter.py \
--nusc_version v1.0-trainval \
--data_root $ROOT_PATH_RAW_NUSCENES_DATA \
--save_root $SAVE_PATH_PROCESSED_NUSCENES_DATA \
Folder structure:
$SAVE_ROOT
├── nuscenes/
│ ├── v1.0-trainval/
│ │ ├── cam_all_train/
│ │ │ ├── v0.0.1/
│ │ │ ├── v0.0.2/
│ ├── v1.0-mini/ (Optional)