Skip to content

Latest commit

 

History

History
43 lines (22 loc) · 1.44 KB

TRAINING.md

File metadata and controls

43 lines (22 loc) · 1.44 KB

LaneGNN Training Guide

In order to train the models described in the paper, we need to first download and pre-process the dataset. To do so, please follow the instructions in urbanlanegraph_dataset/DOWNLOAD.md and urbanlanegraph_dataset/PROCESS_DATA.md.

1 Training of Regressor models

We leverage two centerline regression modules in our paper:

  • lane centerline regression (regressing all visible lane centerlines)
  • ego lane centerline regression (regressing the centerline of the ego-agent lane)

The ego lane centerline regression model gets as input both the RGB images and the lane centerline regression model output.

1.1 Training of lane centerline regression model

python train_centerline_regression.py --dataset-root /path/to/raw/cropped/dataset --sdf_version centerlines-sdf

1.2 Training of ego lane centerline regression model

python train_centerline_regression.py --dataset-root /path/to/raw/cropped/dataset --sdf_version centerlines-sdf-ego-context --checkpoint_path_context_regression /path/to/context_regression_model.pth

Training of LaneGNN model (with pre-trained lane regressors)

python train_lanegnn.py --config lanegnn/config/config.yaml --train_path /path/to/processsed/cropped/dataset/train --eval_path /path/to/processsed/cropped/dataset/eval

The config file contains the following parameters: