Bayesian Continual Learning for Prior Knowledge Integration

Framework for integrating prior knowledge into trajectory prediction models via Bayesian Continual Learning.

Getting started

Installing

1. Clone this repository
2. Download NuScenes Trajectory Prediction Dataset
3. Download Trajectory Data and put into data directory
4. Install Dependencies
   1. pip install requirements.txt
   2. pip install requirements_no_deps.txt --no-deps
   3. For GPU Use: cudatoolkit=11.2.2, cudatoolkit-dev=11.2.2, cudnn=8.2.1.32

Running

1. Run with bayes_covernet/main.py as main entrypoint
2. Edit config.gin to match your needs/setup.
3. First start should use -c and -p for populating the cache. See --help for more arguments and description.

For additional arguments, see python main.py --help. Running in HEBO optimization mode (-o3) requires an additional pip install HEBO==0.3.2

Experiments

Configuration files for reruning experiments can be found under bayes_covernet\config. Please see the paper for hyperparameter configurations. GVCL and Transfer runs require setting a knowledge integrated model checkpoint. GVCL uses a VI_Prior model, set as Trainer.prior_model. Transfer uses a Det_Pretrain model, set as Trainer.load_model.

Model

The provided software computes a trajectory prediction model under consideration of a prior distribution. Currently available are:

Prior Knowledge

• Discrete Driveable: Discrete set of trajectories, as defined by CoverNet, conditioned on all keypoints beeing driveable, as defined by the dataset.

Prediction Model

• Deterministic CoverNet
• Mean-Field Variational CoverNet

Knowledge Integration

• Generalized Variational Continual Learning
• Transfer Learning
• Multitask Learning

Documentation

For further help, see the API-documentation or contact the maintainers.

License

Copyright (c) 2022 Continental Corporation. All rights reserved.

This repository is licensed unter the BSD 3-Clause "New" or "Revised" license. See LICENSE.txt for the full license text.

Changelog

See CHANGELOG.md.