New_approach (suggest name)

Directory Structure:

    alt_cuda_corr
    core
    datasets
    -- config
    -- data_generation
    envs
    general_file
    pretrained_models
    README.md
    runs
    utils

download pretrained_models

link: https://drive.google.com/drive/folders/1PuXwJSyMlG_eYOCDbqB6hU9otF-oK5-4?usp=share_link
directory structure:

pretrained_models
    -- gmflownet_mix-sintel.pth
    -- gmflownet_mix-things.pth
    -- gmflownet-kitti.pth
    -- gmflownet-thing.pth
    -- new_model.pth

dataset generation

carla setup instructions(Debian CARLA installation) The Debain package is available for both Ubuntu 18.04 and Ubuntu 20.04, however the officially supported platform is Ubuntu 18.04.

Set up the Debian repository in the system:

    sudo apt-key adv --keyserver keyserver.ubuntu.com --recv-keys 1AF1527DE64CB8D9
    sudo add-apt-repository "deb [arch=amd64] http://dist.carla.org/carla $(lsb_release -sc) main"

Install CARLA and check for the installation in the /opt/ folder:

    sudo apt-get update # Update the Debian package index
    sudo apt-get install carla-simulator=0.9.12 # Install CARLA

Data generation:
1. Run the CARLA simulator:
```
    CUDA_VISIBLE_DEVICES=0 bash /opt/carla-simulator/CarlaUE4.sh -RenderOffScreen
```
1. generate traffic edit path from file generate_traffic.py on line number 116 i.e
```
client.start_recorder('/home/sushlok/new_approach/datasets/data_generation/recording02.log')
```
with your absolute path and run the following commands
```
    python3 datasets/data_generation/generate_traffic.py
```
run this program for 2-3 mins. it will generate a recording file with all the agents in the scene.
1. append the same path in data_generation.py file on line number 81 i.e
```
client.replay_file("/home/sushlok/new_approach/datasets/data_generation/recording02.log",0,1000,0)
```
edit the path with your absolute path and run the following command
```
    python3 datasets/data_generation/data_generation.py
```
```
4. weather conditions can be changed by commenting and uncommenting the following list from data_generation.py code
```
weather_list = [ # carla.WeatherParameters.SoftRainNight, # carla.WeatherParameters.ClearNoon, carla.WeatherParameters.CloudyNoon, # carla.WeatherParameters.HardRainNight, # carla.WeatherParameters.WetNoon, # carla.WeatherParameters.WetCloudyNoon, # carla.WeatherParameters.MidRainyNoon, # carla.WeatherParameters.HardRainNoon, # carla.WeatherParameters.SoftRainNoon, carla.WeatherParameters.ClearSunset, carla.WeatherParameters.CloudySunset, # carla.WeatherParameters.WetSunset, # carla.WeatherParameters.WetCloudySunset, # carla.WeatherParameters.MidRainSunset, # carla.WeatherParameters.HardRainSunset, # carla.WeatherParameters.SoftRainSunset, # carla.WeatherParameters.ClearNight, # carla.WeatherParameters.CloudyNight, # carla.WeatherParameters.WetNight, # carla.WeatherParameters.WetCloudyNight

]

weather_name_list = [ # 'SoftRainNight', # 'ClearNoon', 'CloudyNoon', # 'HardRainNight', # 'WetNoon', # 'WetCloudyNoon', # 'MidRainyNoon', # 'HardRainNoon', # 'SoftRainNoon', 'ClearSunset', 'CloudySunset', # 'WetSunset', # "WetCloudySunset", # "MidRainSunset", # 'HardRainSunset', # 'SoftRainSunset', # 'ClearNight', # 'CloudyNight', # 'WetNight', # 'WetCloudyNight' ]
final directory structure:

datasets
    -- carla
    -- config
    -- data_generation

create conda environment

    conda env create -f envs/gmflownet.yml

update configuration file with absolute paths and run type in config.yml file and run the training code

    python3 train.py

Dataset generation using Docker

Build the docker container. Navigate to this git repo and run the following command. If you update the source code of this repo and want to use it in the docker container you need to rebuild the docker container, as this repo is copied into the docker container as the last building step:

    docker build -t user/carla_data_generation:TAG_NUMBER .
    docker build -t sushlok/carla_data_generation:0 . # example

Create a docker volume to store the dataset. Run the following command:

    docker volume create --name carla_dataset --opt type=none --opt device=/path/to/local/host/dataset/folder --opt o=bind

ATTENTION! /path/to/local/host/dataset/folder has to belong to the same user ID as the carla user of the docker container which is 1000.

Inside the terminal running the docker container, run the following commands:

    xhost +

    docker run --name carlaserver --mount source=carla_dataset,target=/path/to/docker/dataset/folder -v /tmp/.X11-unix:/tmp/.X11-unix -it --rm --gpus all -p 2000-2002:2000-2002 sushlok/carla_data_generation:0 ./CarlaUE4.sh -RenderOffScreen

Run the generate_traffic and data_generation scripts inside the docker container using the regular carla user. Enter the docker container and run the following commands (make sure not to overwrite previous recordings):

    docker exec -it carlaserver bash
    # Inside the docker container
    cd new_approach/datasets/data_generation
    python3 generate_traffic.py --log_path /path/to/local/docker/dataset/folder/recording02.log
    python3 data_generation.py --log_path /path/to/local/docker/dataset/folder/recording02.log --data_dir /path/to/local/docker/dataset/folder --env_state env_config.yaml

Training using vscode devcontainer

Install docker vscode extension. (first install docker and nvidia-docker on system before downloading vscode extension)
.devcontainer folder contains the docker file and configuration file.
The post processing task is included in the devcontainer.json file which will execute install_requirements.sh file.
After building the enviroment activate gmflownet env by using

conda activate gmflownet

Edit paths from config file related to dataset
Set training procedure, for ex. train: true
finally run train.py code using

python train.py