Skip to content

Latest commit

 

History

History
49 lines (41 loc) · 2.86 KB

README.md

File metadata and controls

49 lines (41 loc) · 2.86 KB
Raw

Hybrid prediction training

This is the code to train and evaluate the models in "HYPER: Learned Hybrid Trajectory Prediction via Factored Inference and Adaptive Sampling".

Tested for Argoverse dataset only.

Package Installation

See the top level README to set up the development environment.

Convert raw Argoverse data to protobufs

See instructions in here

Train a model

Example training script for Argoverse:

python intent/multiagents/hybrid/train_hybrid_vehicle_prediction.py --input-training-dir ~/intent/argoverse_v11/train_pb/ --input-validation-dir ~/intent/argoverse_v11/val_pb/ --dropout 0.1 --child-network-dropout 0.0 --additional-dropout-ratio 0.2 --training-set-ratio 0.8 --batch-size 32 --epoch-size 512 --learn-discrete-proposal --map-points-max 3600 --map-encoder-type attention --report-sample-metrics true --trajectory-regularization-cost 1.0

Options for baselines and ablations:

  • --learn-discrete-proposal: Remove this option will use the transition function instead of proposal function for discrete prediction.
  • --proposal-adaptive-sampling: Set to false to use the non-adapative sampling method.
  • --hybrid-fixed-mode: Add this option to disable mode transitions.

Evaluate model offline

Given a trained model with saved model directory name SESSION_ID, run

python intent/multiagents/hybrid/run_hybrid_vehicle_trajectory_prediction.py --input-training-dir ~/intent/argoverse_v11/train_pb/ --input-validation-dir ~/intent/argoverse_v11/val_pb/ --dropout 0.1 --child-network-dropout 0.0 --additional-dropout-ratio 0.2 --training-set-ratio 0.8 --batch-size 32 --epoch-size 512 --learn-discrete-proposal--map-points-max 3600 --map-encoder-type attention --resume-session-name SESSION_ID_best_fde --MoN-number-samples 50 --hybrid-runner-subsample-size 6 --hybrid-runner-save True

Options include:

  • --hybrid-runner-subsample-size: Number of samples to select for final prediction. Default is 6.
  • --hybrid-runner-nms-dist-threshold: Distance threshold used for NMS. Default is 2.0.
  • --hybrid-runner-dist-type: Distance type used for FPS and NMS. Option includes 'final' (default) and 'avg'.
  • --hybrid-runner-visualize: Whether to visualize examples.
  • --hybrid-runner-save: Whether to save individual prediction results into json files.

Unit Test

Test basic training with example Argoverse data:

python intent/multiagents/hybrid/test_hybrid_training.py

Citation

@inproceedings{huang2022:hyper,
  title={HYPER: Learned Hybrid Trajectory Prediction via Factored Inference and Adaptive Sampling},
  author={Huang, Xin and Rosman, Guy and Gilitschenski, Igor and Jasour, Ashkan and McGill, Stephen G. and Leonard, John J. and Williams, Brian C.},
  booktitle={2022 IEEE International Conference on Robotics and Automation (ICRA)},
  year={2022}
}