- Serve as a partial reimplementation of the ESP multi-agent forecasting model
- Train a Deep Imitative Model for CARLA autonomous driving task. See https://github.com/nrhine1/deep_imitative_models.
precog/esp_train.py
Interface to train a model
precog/esp_infer.py
Interface to perform test-time inference (plotting and metrics computation)
export PRECOGCONDAENV=pre3
conda create -n $PRECOGCONDAENV python=3.6.6
conda activate $PRECOGCONDAENV
source precog_env.sh
pip install -r requirements.txt
Potentially install the nrhine1/deep_imitative_models
repo.
export CUDA_VISIBLE_DEVICES=0; python $PRECOGROOT/precog/esp_train.py \
bijection=social_convrnn \
dataset=trimodal_dataset \
bijection.params.A=1 \
dataset.params.B=20 \
main.eager=false
export CUDA_VISIBLE_DEVICES=0; python $PRECOGROOT/precog/esp_train.py \
dataset=carla_town01_A1_T20_v2 \
main.eager=False \
bijection.params.A=1 \
optimizer.params.plot_before_train=True \
optimizer.params.save_before_train=True
Download the nuscenes dataset, then use the script preprocess_nuscenes.py
Recall that the log-likelihood is insensitive to sample quality. If you're not using a sample-penalizing metric, it will take longer training time to observe higher-quality samples.
@InProceedings{Rhinehart_2019_ICCV,
author = {Rhinehart, Nicholas and McAllister, Rowan and Kitani, Kris and Levine, Sergey},
title = {PRECOG: PREdiction Conditioned on Goals in Visual Multi-Agent Settings},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}