Skip to content

README.md

Latest commit

 

History

History
45 lines (34 loc) · 1.88 KB

README.md

File metadata and controls

45 lines (34 loc) · 1.88 KB
Raw

Training on Savio

The job files use the ic_piml account on savio4_htc (billed per core). Change them if you use another account.

First time

module load python/3.11.6-gcc-11.4.0
cd /global/scratch/users/$USER
git clone https://github.com/clementdumeril/Fixed-Wing.git && cd Fixed-Wing/python
python -m venv ~/.venvs/fixed-wing && source ~/.venvs/fixed-wing/bin/activate
pip install torch --index-url https://download.pytorch.org/whl/cpu   # CPU build, much smaller
pip install -e ".[rl]"
mkdir -p runs

Every session

cd /global/scratch/users/$USER/Fixed-Wing/python
module load python/3.11.6-gcc-11.4.0 && source ~/.venvs/fixed-wing/bin/activate
git pull   # if job files were edited by hand before: git checkout -- savio/ first

Runs

  • Smoke test, a few minutes: sbatch savio/smoke_test.slurm
  • One training: sbatch --array=0 savio/train_sac.slurm; seeds 1 to 4: sbatch --array=1-4 savio/train_sac.slurm
  • Flight conditions, passed to train_sac.py: sbatch --array=0-2 savio/train_sac.slurm --wind-max 5 --turbulence light --noise 1
  • Follow: squeue -u $USER, grep ep_len_mean runs/slurm-<job>_0.out | tail
  • Result: cat runs/sac_<job>_seed0/evaluation.json, same metrics and test scenarios as examples/pid_baseline.py

Jobs keep running after you log out. Each run writes runs/sac_<job>_seed<k>/: progress.csv, checkpoints/, model.zip, evaluation.json.

Plots, on your computer

Copy the runs back (from your computer, through the data transfer node), then:

scp -r <user>@dtn.brc.berkeley.edu:/global/scratch/users/<user>/Fixed-Wing/python/runs .
python examples/plot_training.py runs/sac_<job>_seed0 --out training.png
python examples/compare_flights.py --model runs/sac_<job>_seed0/model.zip --seeds 0 1 2 --out flights.png
python examples/ui_policy.py --model runs/sac_<job>_seed0/model.zip --seed 0   # 3D, with ui-server running