Download and prepare nuscenes dataset following UniAD's instruction to ./UniAD/data
Step 1: create a deployment docker container and run
docker run -it --gpus all --shm-size=8g -v </host/system/path/to/UniAD>:/workspace/UniAD uniad_torch1.12 /bin/bash
Step 2: inside the deployment docker container, build uniad_mmdet3d
cd /workspace/UniAD/third_party/uniad_mmdet3d/
python3 setup.py build develop --user
Inside deployment docker container, generate six inputs to ./UniAD/nuscenes_np/uniad_onnx_input
for ONNX exportation, and NUM_FRAME
preprocessed inputs to ./UniAD/nuscenes_np/uniad_trt_input
for inference application. By default we set NUM_FRAME
to 69
which covers the first two scenes, user can choose any number in the range of [6, 6018]
.
cd /workspace/UniAD
PYTHONPATH=$(pwd) python3 ./tools/process_metadata.py --num_frame NUM_FRAME
<- Last Page: Environment Preparation
-> Next Page: UniAD-tiny Traning and Exportation