Preparation of GoalFlow Environment

Follow the steps below to set up the GoalFlow environment.

1. Clone the GoalFlow Repository

First, clone the repository and create a data directory:

git clone https://github.com/YvanYin/GoalFlow.git
cd GoalFlow
mkdir data

Make sure that the navsim_log_path and sensor_blobs_path in default_evaluation.yaml and default_metric_caching.yaml match your dataset path.

2. Install packages

Firstly, create conda environment with python 3.10 and intall the packages

conda create -n goalflow python=3.10
conda activate goalflow
pip install -r requirements.txt
pip install -e nuplan-devkit
pip install -e .

2. Prepare the Cache

In NAVSIM, it is recommended to store features and metrics as a cache to speed up training and evaluation.

Step 1: Cache Features

Cache features to train the model and generate trajectories.

• If you only need to evaluate the model, you only need to run run_dataset_cache_test.sh.

sh scripts/cache/run_dataset_cache_test.sh
sh scripts/cache/run_dataset_cache_trainval.sh

Step 2: Cache Metrics

Cache metrics to evaluate trajectories. The test cache is used to generate DAC score labels for training the goal point module.

• If you only need to evaluate your model, you only need to run run_metric_caching_test.sh.

sh scripts/cache/run_metric_caching_test.sh
sh scripts/cache/run_metric_caching_trainval.sh

Step 2: Cache DAC score labels (Optional)

Cache DAC score labels. You need to specify METRIC_PATH obtained from run_metric_caching_trainval.sh

• We also provide precomputed DAC score labels, which you can directly use.

sh scripts/generate/run_generate_dac_label.sh