def main():
args = parse_args()
# Hyperparameters
params = hpo.Params([
['--hidden-dim', 32, [16, 32, 64, 128, 256]],
['--n-edge-layers', 4, [1, 2, 4, 8]],
['--n-node-layers', 4, [1, 2, 4, 8]],
['--weight-decay', 1.e-4, (0., 1e-3)],
['--n-graph-iters', 8, (4, 16)],
['--real-weight', 3., (1., 6.)],
['--lr', 0.001, [1e-5, 1e-4, 1e-3, 1e-2]],
])
# Define the command to be run by the evaluator
output_dir = "'${SCRATCH}/heptrkx/results/hpo_%s_${SLURM_JOB_ID}_${SLURM_STEP_ID}'" % (
os.path.basename(args.config).split('.')[0])
cmd = ('python -u train.py %s' % args.config +
' --rank-gpu -d ddp-file --fom best' +
' --n-epochs %i' % args.epochs + ' --output-dir %s' % output_dir)
# SLURM options
n_nodes = (args.nodes if args.nodes is not None else int(
os.environ['SLURM_JOB_NUM_NODES']))
n_tasks_per_node = (args.ntasks_per_node
if args.ntasks_per_node is not None else int(
os.environ['SLURM_NTASKS_PER_NODE']))
n_tasks_per_eval = args.nodes_per_eval * n_tasks_per_node
alloc_args = ('-J hpo %s --time %s' % (args.alloc_args, args.time) +
' --ntasks-per-node %i' % n_tasks_per_node)
launch_args = '-n %i -u' % n_tasks_per_eval
# Define the evaluator
evaluator = hpo.Evaluator(cmd,
nodes=n_nodes,
workload_manager='slurm',
alloc_args=alloc_args,
launch_args=launch_args,
nodes_per_eval=args.nodes_per_eval,
verbose=True)
# Random search optimizer
if args.alg == 'random':
optimizer = hpo.RandomOptimizer(evaluator, num_iters=args.iters)
# Genetic search optimizer
else:
results_file = 'hpo.log'
optimizer = hpo.GeneticOptimizer(evaluator,
pop_size=args.pop_size,
num_demes=args.demes,
generations=args.generations,
mutation_rate=args.mutation_rate,
crossover_rate=args.crossover_rate,
verbose=True,
log_fn=results_file)
# Run the Optimizer
optimizer.optimize(params)
def main():
args = parse_args()
# Set up evaluator
# configs/hpo_cifar10_cnn.yaml is a scaled down version of cifar10
# --hpo is required for train.py to print the FoM
# --no-output is required to avoid checkpointing
eval_cmd = 'python ./train.py configs/hpo_cifar10_cnn.yaml --hpo --no-output'
evaluator = hpo.Evaluator(eval_cmd,
nodes=args.num_nodes,
verbose=args.verbose)
# Set up search space for HPs: learning rate and dropout
params = hpo.Params([['--optimizer lr', 0.001, (1e-6, 1)],
['--dropout', 0.1, (0.0, 0.5)]])
# Set up genetic optimizer with 8 evaluations/generation and 3 generations
optimizer = hpo.GeneticOptimizer(evaluator,
generations=4,
pop_size=8,
num_demes=1,
mutation_rate=0.6,
verbose=args.verbose)
# Optimize the hyperparameters
optimizer.optimize(params)
# Print the figure of merit value for the best set of hyperparameters
print(optimizer.best_fom)
# Print the best set of hyperparameters found
print(optimizer.best_params)
def main():
args = parse_args()
# Hardcode some config
#n_nodes = 4 #32
#config_file = 'configs/test.yaml'
#pop_size = 2 #16
#n_demes = 2 #4
#n_generations = 4
#mutation_rate = 0.05
#crossover_rate = 0.33
#alloc_args='-J hpo -C haswell -q interactive -t 4:00:00'
#checkpoint_dir = 'checkpoints'
# Hyperparameters
params = hpo.Params([['--lr', 0.001, (1e-6, 0.1)],
['--n-graph-iters', 4, (1, 16)],
['--real-weight', 3., (1., 6.)]])
# Define the command to be run by the evaluator
cmd = 'python train.py %s' % args.config
cmd += ' --fom last --n-epochs 1 --resume --output-dir @checkpoint'
# Define the evaluator
result_dir = os.path.expandvars('$SCRATCH/heptrkx/results/pbt_%s' %
time.strftime('%Y%m%d_%H%M%S'))
evaluator = hpo.Evaluator(cmd,
run_path=result_dir,
nodes=args.nodes,
launcher='wlm',
verbose=True,
nodes_per_eval=args.nodes_per_eval,
checkpoint='checkpoints',
alloc_args=args.alloc_args)
# Define the Optimizer
optimizer = hpo.GeneticOptimizer(evaluator,
pop_size=args.pop_size,
num_demes=args.demes,
generations=args.generations,
mutation_rate=args.mutation_rate,
crossover_rate=args.crossover_rate,
verbose=True)
# Run the Optimizer
optimizer.optimize(params)
argparser = argparse.ArgumentParser()
argparser.add_argument('-N', '--num_nodes', type=int, default=1)
argparser.add_argument('--generations', type=int, default=3)
argparser.add_argument('--num_demes', type=int, default=1)
argparser.add_argument('--pop_size', type=int, default=8)
argparser.add_argument('--mutation_rate', type=float, default=0.5)
argparser.add_argument('--crossover_rate', type=float, default=0.33)
argparser.add_argument('--verbose', action='store_true')
args = argparser.parse_args()
print("------------------------------------------------------------")
print("Genetic HPO Example: LeNet-5 (MNIST) TensorFlow -- Cray Inc.")
print("------------------------------------------------------------")
evaluator = hpo.Evaluator('python3 {0}/mnist.py'.format(src_path),
run_path=run_path,
src_path=src_path,
nodes=args.num_nodes)
optimizer = hpo.GeneticOptimizer(evaluator,
generations=args.generations,
num_demes=args.num_demes,
pop_size=args.pop_size,
mutation_rate=args.mutation_rate,
crossover_rate=args.crossover_rate,
verbose=args.verbose,
log_fn='mnist-topology.log')
params = hpo.Params([["--dropout", 0.5, (0.005, 0.9)],
["--momentum", 1.0e-4, (1.0e-6, 1.0e-2)],
["--c1_sz", 5, (2, 8)], ["--c1_ft", 32, (8, 128)],
["--c2_sz", 5, (2, 8)], ["--c2_ft", 64, (16, 256)],
from crayai import hpo
argparser = argparse.ArgumentParser()
argparser.add_argument('--generations', type=int, default=250)
argparser.add_argument('--num_demes', type=int, default=1)
argparser.add_argument('--pop_size', type=int, default=100)
argparser.add_argument('--mutation_rate', type=float, default=0.05)
argparser.add_argument('--crossover_rate', type=float, default=0.33)
argparser.add_argument('--verbose', action='store_true')
args = argparser.parse_args()
print("------------------------------------------------------------")
print("Genetic HPO Example: LeNet-5 (MNIST) TensorFlow -- Cray Inc.")
print("------------------------------------------------------------")
evaluator = hpo.Evaluator('python3 source/mnist.py', run_path='runPath')
optimizer = hpo.genetic.Optimizer(evaluator,
generations=args.generations,
num_demes=args.num_demes,
pop_size=args.pop_size,
mutation_rate=args.mutation_rate,
crossover_rate=args.crossover_rate,
verbose=args.verbose,
log_fn='mnist-topology.log')
params = hpo.Params([["--dropout", 0.5, (0.005, 0.9)],
["--momentum", 1.0e-4, (1.0e-6, 1.0e-2)],
["--c1_sz", 5, (2, 8)], ["--c1_ft", 32, (8, 128)],
["--c2_sz", 5, (2, 8)], ["--c2_ft", 64, (16, 256)],
["--fullyc_sz", 1024, (64, 4096)]])
#!/usr/bin/env python3
# encoding: utf-8
"""Random optimizer example"""
from crayai import hpo
evaluator = hpo.Evaluator('python source/sin.py', verbose=True)
params = hpo.Params([["-a", 1.0, (-1.0, 1.0)], ["-b", -1.0, (-1.0, 1.0)],
["-c", 1.0, (-1.0, 1.0)], ["-d", -1.0, (-1.0, 1.0)],
["-e", 1.0, (-1.0, 1.0)], ["-f", -1.0, (-1.0, 1.0)],
["-g", 1.0, (-1.0, 1.0)]])
optimizer = hpo.RandomOptimizer(evaluator, num_iters=129, verbose=True)
optimizer.optimize(params)
print(optimizer.best_fom)
print(optimizer.best_params)
minimize both error and time taken when fitting polynomial to sin function.
"""
from crayai import hpo
import argparse
parser = argparse.ArgumentParser(description=__doc__,
formatter_class=argparse.RawTextHelpFormatter)
parser.add_argument('--source',
type=str,
default='sin-time-accuracy.py',
help='source script')
args = parser.parse_args()
evaluator = hpo.Evaluator('python source/{0}'.format(args.source))
params = hpo.Params([["-a", 1.0, (-10.0, 10.0)], ["-b", -1.0, (-10.0, 10.0)],
["-c", 1.0, (-10.0, 10.0)], ["-d", -1.0, (-10.0, 10.0)],
["-e", 1.0, (-10.0, 10.0)], ["-f", -1.0, (-10.0, 10.0)],
["-g", 1.0, (-10.0, 10.0)]])
optimizer = hpo.genetic.Optimizer(evaluator,
verbose=True,
generations=5,
pop_size=10,
log_fn='genetic.log')
optimizer.optimize(params)
print(optimizer.best_fom)
["--lr", 0.001, (1e-6, 0.1)],
["--real-weight", 3, (1, 5)]])
config_file = 'configs/agnn.yaml'
## Define Command to be run by the evaluator
cmd = "python train.py %s --crayai --resume --ranks-per-node 1 -v --gpu 0" % config_file
if args.run_pbt:
cmd += " --pbt_checkpoint @checkpoint"
## Define the evaluator
timestr = time.strftime("%Y%m%d-%H%M%S")
evaluator = hpo.Evaluator(
cmd,
run_path="./runs/run%s" % timestr,
nodes=args.nodes,
launcher='wlm',
verbose=True,
nodes_per_eval=args.nodes_per_eval,
checkpoint=args.checkpoint if args.run_pbt else '',
alloc_args="-J agnn-heptrkx --exclusive --time=24:00:00 -C P100 --gres=gpu",
alloc_jobid=args.alloc_job_id)
## Define the Optimizer
optimizer = hpo.genetic.Optimizer(evaluator,
gens_per_epoch=args.gens_per_epoch,
pop_size=args.pop_size,
num_demes=1,
generations=args.generations,
mutation_rate=args.mutation_rate,
crossover_rate=args.crossover_rate,
verbose=True)