Inference using UniAD

This repository is a fork of the UniAD, used to showcase how to implment a model to be evaluated according to the NeuroNCAP evaluation framework.

Changes

This repository differs from the original UniAD repository in the following ways:

• Added a config file at projects/configs/stage2_e2e/inference_e2e.py to limit the operations applied to the input.
• Added inference functionality in the inference folder. This includes two files:
  • runner.py which wraps the original UniAD model to be able to run in inference mode (original can only be ran in training or validation/testing mode).
  • server.py which is a simple FastAPI server that opens endpoints to run inference using the model. The endpoints follow the NeuroNCAP API specification.
• Added a Dockerfile that was used to build the .sif file that the model can run in.

How to use

1. Download the weights:

mkdir checkpoints
wget "https://github.com/OpenDriveLab/UniAD/releases/download/v1.0.1/uniad_base_e2e.pth" -P checkpoints
wget https://github.com/OpenDriveLab/UniAD/releases/download/v1.0/motion_anchor_infos_mode6.pkl -P checkpoints

2. Build the .sif file:

docker build -t uniad:latest -f docker/Dockerfile .
singularity build uniad.sif docker-daemon://uniad:latest

Links:

• How to install docker
• How to install singularity

3. Follow the instructions in the NeuroNCAP repository.

Original UniAD README.md