This is the code accompanying the lecture on self-driving with carla that can be found here https://www.youtube.com/watch?v=MNiqlHC6Kn4&t You should only run this after installing carla and getting a grasp on how to run the simulator and calling the python API, which is documented very well on their site.
I did not create a dependency or yml file (will do so at a later time), but you need carla, pygame, pytorch, opencv and numpy to run this project
You should ensure that you have a weights folder when you run the project. If you do not have one, then just run initial_setup.py and it will create it for you. If you just cloned the repository, I reccomend you run this file first.
Run this file if you want to evaluate the performance of your agent
env = SimEnv(visuals=False)
The call above initializes our simulation environment. You should set visuals to False if you do not want to open this with pygame, or to True if you want a pygame window to open along with the simulator.
model.load('weights/model_ep_4400')
This loads a trained/pre-trained model. The program will not run unless it can load this model. The 4400 indicates that this model was trained for 4400 episodes. For example, if you train your own model for 200 episodes you will see the following files in the weights folder
model_ep_200_optimizer and model_ep_200_Q
You can then load the model as model.load('weights/model_ep_200'). Please note however that this is likely to be a very bad model, and it will learn effectively after many episodes.
This is for training the model. The model only starts learning after a certain number of episodes, and it can take from 8-10 hours (at least on my setup) before we see signs of learning. I will now describe a few variables you can set to configure your training process. You can modify them yourself in config.py.
target_speed --> Speed you want the car to move at in km/h
max_iter --> Maximum number of steps before starting a new episode
start_buffer --> Number of episodes to run before starting training
train_freq --> How often to train (set to 1 to train every step, 2 to train every 2 steps etc)
save_freq: --> Frequency of saving our model
start_ep --> Which episode should we start on (just a counter which you can update if program crushes while training for example)
max_dist_from_waypoint --> Maximum distance from waypoint/road before we decide to terminate the episode