Note that, except for deadly corridor and death match, A3C has gained not bad results on the rest scenarios
In every scenario, there is a network.py script, an agent.py script and a main.py file to run the code
All the configurations are in the configs.py file in each scenario. Or you can use the shell scripts to run it.
The contents in both configs.py and shell scripts are self-explained.
The check_point directory stores checkpoint data of the neural network in each specific step, and the summaries
directory stores events files contain data while helps developers to understand what's going on with the model only during training stage.
You can use tensorboard command to visualise the data, for example
setsid tensorboard --logdir=D3_battle --port=7400 > logs.D3_battle 2>&1Note that, if you want to train the model, no GUI will be prompted out since game window will not visible as you can see
in the code game.set_window_visible(self.play), when training self.play is False.
So, after training a model, how can I run it locally? You should set IS_TRAIN in config.py to False and set the model_file value, for instance 'model-30150.ckpt'.
To get started, you can run the code of the code of healthpack gathering scenario to get familiar with the code.
If you have any problem, please open an issue!
The code of A3C framework was modified from awjuliani's repo.
Try setsid command to run the program in background, for example
setsid python3 main.py > train.log 2>&1
If you want to kill a running job on the server, use pkill -TERM -P THE_PID to kill all its children processes in stead of kill -9 THE_PID.
Get the runing processes' directory
$ for id in `ps -ef | grep main.py | awk '{print $2}'`; do ls -l /proc/${id}/cwd ; done
ls: cannot access /proc/10800/cwd: No such file or directory
lrwxrwxrwx 1 doom doom 0 Aug 30 23:19 /proc/23963/cwd -> /home/doom/Code/battle_new
lrwxrwxrwx 1 doom doom 0 Aug 29 19:30 /proc/38867/cwd -> /home/doom/Code/battle
lrwxrwxrwx 1 doom doom 0 Sep 2 00:10 /proc/39261/cwd -> /home/doom/Code/battle_decay
lrwxrwxrwx 1 doom doom 0 Sep 2 00:10 /proc/39706/cwd -> /home/doom/Code/battle_new_2Get the utilization of GPU for every 5 seconds
while true; do nvidia-smi --query-gpu=utilization.memory --format=csv && nvidia-smi --query-gpu=utilization.gpu --format=csv; sleep 5; doneYou can watch the videos of the results on Youtube.