Skip to content

Latest commit

 

History

History
executable file
·
44 lines (37 loc) · 1.5 KB

preparation.md

File metadata and controls

executable file
·
44 lines (37 loc) · 1.5 KB

Prepare nuScenes-Occupancy

The data processing in DriveDreamer-2 is the same as that in DriveDreamer. Additionally, to avoid redundant calculations of identical prompts during training, we have precomputed and stored the embeddings for these prompts in HERE.

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)