MMFN: Multi-Modal Fusion Net for End-to-End Autonomous Driving

official branch for IROS 2022 paper codes including collecting data, all benchmark in paper, training scripts and evaluations etc.

Quickly view

Background: How to efficiently use high-level information or sensor data in end-to-end driving. The whole architecture in MMFN from origin paper:

**Scripts** quick view in `run_steps` folder:

• phase0_run_eval.py : collect data/ run eval in select routes files or town map
• phase1_preprocess.py : pre-process data before training to speed up the whole training time
• phase2_train.py: after having training data, run this one to have model parameters files. pls remember to check the config (You can try training process on Docker also)

This repo also provide experts with multi-road consideration. Refer to more experts, pls check latest carla-expert repo

0. Setup

Install anaconda

wget https://repo.anaconda.com/archive/Anaconda3-2020.11-Linux-x86_64.sh
bash Anaconda3-2020.11-Linux-x86_64.sh
source ~/.profile

Clone the repo and build the environment

git clone https://github.com/Kin-Zhang/mmfn
cd mmfn
conda create -n mmfn python=3.7
conda activate mmfn
pip3 install -r requirements.txt

For people who don't have CARLA [在内地的同学可以打开scripts换一下函数 走镜像下载更快点.. ]

chmod +x scripts/*
./scripts/setup_carla.sh
# input version
10.1
# auto download now ...

Download the vectorized lib with .so, will install based on your system version, for commercial reason, we cannot open this part of codes(C++) fully but get the lib file, with lib setting, we can still directly input the opendrive file and output the vectorized lane, pls see more usage in the code

./scripts/setup_lib.sh

bashrc or zshrc setting:

# << Leaderboard setting
# ===> pls remeber to change this one
export CODE_FOLDER=/home/kin/mmfn
export CARLA_ROOT=/home/kin/CARLA_0.9.10.1
# ===> pls remeber to change this one
export SCENARIO_RUNNER_ROOT=${CODE_FOLDER}/scenario_runner
export LEADERBOARD_ROOT=${CODE_FOLDER}/leaderboard
export PYTHONPATH="${CARLA_ROOT}/PythonAPI/carla/":"${SCENARIO_RUNNER_ROOT}":"${LEADERBOARD_ROOT}":"${CARLA_ROOT}/PythonAPI/carla/dist/carla-0.9.10-py3.7-linux-x86_64.egg":"${CODE_FOLDER}/team_code":${PYTHONPATH}

1. Dataset

The data is generated with leaderboard/expert_agent/mmfn_pilot.py in 8 CARLA towns using the routes and scenarios files provided at leaderboard/data on CARLA 0.9.10.1

The dataset is structured as follows:

- TownX_{tiny,short,long}: corresponding to different towns and routes files
    - TownX_X_07_01_23_26_34: contains data for an individual route
        - rgb_front: multi-view camera images at 400x300 resolution
        - lidar: 3d point cloud in .npy format
        - measurements: contains ego-agent's position, velocity and other metadata
        - radar: radar point in .npy format
        - maps: birdview maps image
        - opendrive: opendrive xodr file

How to Generate

First, please modify the config files and on .zshrc or .bashrc remember to export your CARLA_ROOT

# please remember to change this!!! TODO or will change by modi
absolute_path: '/home/kin/mmfn'
carla_sh_path: '/home/kin/CARLA_0.9.10.1/CarlaUE4.sh'

# Seed used by the TrafficManager (default: 0)
port: 2000
trafficManagerSeed: 0

# ============== for all route test=============== #
debug: False

# only for debug ===> or just test the agent
routes: 'leaderboard/data/only_one_town.xml'
# towns: ['Town01', 'Town02', 'Town06', 'Town07']
# routes: 'leaderboard/data/training_routes/'

scenarios: 'leaderboard/data/all_towns_traffic_scenarios.json'

# ====================== Agent ========================= #
track: 'MAP' # 'SENSORS'
agent: 'team_code/expert_agent/auto_pilot.py'
defaults:
  - agent_config: expert

Please write great port according to the CARLA server, and inside the script it will try to use Epic or vulkan mode since opengl mode will have black point on raining day, after setting the absolute_path and carla_sh_path, directly run this: [Hine: If the GPU memeory is bigger than 12G, and CPU memory with big swap space will be better.]

conda activate mmfn
python run_steps/phase0_run_eval.py port=2010 towns="['Town02', 'Town04', 'Town05', 'Town10']" resume=True if_open_carla=True

# if you have big GPU memory, you can start with these two simultaneously.
python run_steps/phase0_run_eval.py port=2000 towns="['Town01', 'Town03', 'Town06', 'Town07']" resume=True if_open_carla=True

The dataset folder tree will like these one:

data
└── train
	├── Town06_tiny
        ├── Town06_00_07_01_23_40_19
            ├── lidar
            ├── maps
            ├── measurements
            ├── opendrive
            ├── rgb_front
            └── radar
        ├── Town06_00_07_01_23_40_19
        └── Town06_00_07_01_23_40_19

The scripts will start a CARLA with specific port number, example running like this:

Note: since official stable leaderboard have memory leak issue, check out here. Please check your memory and kill it to resume. Other notes for CARLA leaderboard check out here

2. Training

No need CARLA in these phase, please remember to modify the train.yml config file and especially ==modify the DATA PATH==

The record and visualization on training params use the wandb, please login before train, more details can be found at wandb.ai, You can disable wandb from config file using disabled

1. To speed up the training since the data process is somehow time consuming. Save the finished data:

   python run_steps/phase1_preprocess_data.py

2. DDP if you want to use multi-GPU through DDP or multi computers: nproc_per_node as using GPU，nnodes as computer number

   CUDA_VISIBLE_DEVICES=0,1,2,3,4,5 python -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 run_steps/phase2_train_net.py

3. Or one single GPU can just run:

   python run_steps/phase2_train_net.py

Note for Docker: There is a Dockerfile ready for building training environment, but please remember to using -v link the datasets folder to container.

Benchmark

fork from transfuser codes, the benchmark training file and process can run as following command:

python team_code/benchmark/aim/train.py --device 'cuda:0'
python team_code/benchmark/cilrs/train.py --device 'cuda:0'
python team_code/benchmark/transfuser/train.py --device 'cuda:0'

3. Evaluate

This part is for evaluating to result or leaderboard, you can also download the modal file and try upload to leaderboard through leaderbaord branch. The whole process is the same way like generate dataset

1. Download or Train a model file saved to log/mmfn_img or log/mmfn_vec or log/mmfn_rad

2. Keep config/eval.yaml same as collect.yaml but modified model file location as first step side

   routes: 'leaderboard/data/only_one_town.xml'
   # ====================== Agent ========================= #
   track: 'MAP' # 'SENSORS'
   agent: 'team_code/e2e_agent/mmfn_imgnet.py'
   defaults:
     - agent_config: e2e
   
   # ======== e2e.yaml file for model path ============ #
   model_path: 'log/mmfn_img' 

3. Running eval python script and see result json file in result Folder

   python run_steps/phase0_run_eval.py

If all setting is correct, you will see eval like this one, pls remember to change route to only_one_town for debug.

Cite Us

@inproceedings{zhang2022mmfn,
  title={MMFN: Multi-Modal-Fusion-Net for End-to-End Driving},
  author={Zhang, Qingwen and Tang, Mingkai and Geng, Ruoyu and Chen, Feiyi and Xin, Ren and Wang, Lujia},
  booktitle={2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages={8638--8643},
  year={2022},
  organization={IEEE}
}

Acknowledgements

This implementation is based on codes from several repositories. Thanks for these authors who kindly open-sourcing their work to the community. Please see our paper reference part to get more information on our reference

• LBC, WorldOnRails

• Transfuser

• CARLA Leaderboard, Scenario Runner

• carla-brid-view

• pylot