This section outlines the steps and commands that need to be executed to train on a GPU cluster. For our case, the training is carried out on an NVIDIA RTX 3090 GPU.
In order to train with an NVIDIA GPU RTX3090 you need the latest version of pytorch. Inside your rosnav venv, do:
workon rosnav
pip3 install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
The training jobs could run several days, so you should use tmux. This way you can create a virtual session which will keep running on the server, even if the SSH connection is interrupted.
- Install tmux
sudo apt-get update
sudo apt-get install tmux
- You will need two terminals: one for the arena simulation and one for the training script.
-Create screens using tmux :
tmux simulation
tmux training
If you just type tmux, the session name will be a enumerated (0,1,2,...)
- Then run the training commands listed in section 1 below.
-To detach screens during training (for example if you want to power off your own machine), press
CTRL-B and afterwards CTRL-D
-To reatach screens to check progress, run these commands in terminals:
tmux a -t training
Explanation: a: attach to a session -t: flag to indicate name of the session training: name of the session
For more shortcuts, see https://tmuxcheatsheet.com/
The first terminal (the simulation session, if using Screen) is needed to run arena. Run these four commands:
source $HOME/arena_ws/devel/setup.zsh
workon rosnav
roslaunch arena_bringup start_training.launch train_mode:=true use_viz:=false task_mode:=random map_file:=map_small num_envs:=24
A second terminal (the training session if using Screen) is needed to run the training script. Run these four commands:
source $HOME/arena_ws/devel/setup.zsh
workon rosnav
roscd arena_local_planner_drl
Now, run one of the two commands below to start a training session:
python scripts/training/train_agent.py --load pretrained_ppo_mpc --n_envs 24 --eval_log
python scripts/training/train_agent.py --load pretrained_ppo_baseline --n_envs 24 --eval_log
When the training script is done, it will print the following information and then exit:
Time passed: {time in seconds}s
Training script will be terminated
Please make a note of this time taken.
The simulation will not terminate automatically, so it needs to be stopped with CTRL-C.