propagation-code-optimization

Initialize environment

Create a .env file with

export USER= 
export GROUP=

and initialize it with source .env.

Uploading

make upload

Downloading log files

make get_logs

Deployment Examples

python3 -m optimizer.main --algorithm hill_climbing --steps 4
python3 -m optimizer.main --algorithm greedy --steps 4
python3 -m optimizer.main --algorithm tabu_greedy --steps 4 --hparams '{"n_tabu":5}'
python3 -m optimizer.main --algorithm simulated_annealing --steps 10 --hparams '{"t0":20}'
python3 -m optimizer.main --algorithm csa --steps 10 --batch
python3 -m optimizer.main --algorithm cmaes --steps 10

Flag --batch: runs 4 instances. Can either

• Calculate the cost function in parallel (CSA, CMAES)
• Runs sequential programs in different instances in parallel with different initializations (Hill Climbing, Greedy, Tabu Greedy, Simulated Annealing)

Scripts

Test affinity parameters: run

/usr/bin/mpirun -np 1 -map-by ppr:1:node:PE=16 python3 optimizer/test_affinity.py

Hyperparameter optimization:

/usr/bin/mpirun -np 1 -map-by ppr:1:node:PE=16 python3 optimizer/test_sa_hyperparams.py
/usr/bin/mpirun -np 1 -map-by ppr:1:node:PE=16 python3 optimizer/test_sa_hyperparams.py --algorithm csa --steps 50

FAQ

Installation of CMA: Check if the pip used to install is the same as the Python being used to run. Sometimes we need to use /usr/bin/pip (TODO: avoid this). Installation of Nevergrad: for the BayesianOptimization, Nevergrad should be installed with the same instructions as the CMA