ytopt is a machine-learning-based search software package that consists of sampling a small number of input parameter configurations,
evaluating them, and progressively fitting a surrogate model over the input-output space until exhausting the user-defined time or maximum number of
evaluations. The package provides two different class of methods: Bayesian Optimization and Reinforcement Learning.
The software is designed to operate in the master-worker computational paradigm, where one master node fits
the surrogate model and generates promising input configurations and worker nodes perform the computationally expensive evaluations and
return the outputs to the master node.
The asynchronous aspect of the search allows the search to avoid waiting for all the evaluation results before proceeding to the next iteration. As
soon as an evaluation is finished, the data is used to retrain the surrogate model, which is then used to bias the search toward the
promising configurations.
docs/
Sphinx documentation files
ppo/
proximal policy optimization based reinforcement learning
problems/
easy to evalaute benchmark functions
test/
scipts for running benchmark problems in the problems directory
ytopt/
scripts that contain the search implementations
conda create -n ytopt -c anaconda python=3.6
source activate ytopt
git clone https://github.com/ytopt-team/ytopt.git
cd ytopt/
pip install -e .
If you encounter mpi4py installtion error, (re)install mpich as follows
conda install -c conda-forge mpich
An example is given in problems/ackley_mix
To define an autotuning problem, create two files.
The problem.py file defines the search space:
from collections import OrderedDict
import numpy as np
import os
np.random.seed(0)
HERE = os.path.dirname(os.path.abspath(__file__))
from ytopt.problem import Problem
cmd_frmt = "python " +HERE+"/executable.py"
nparam = 6
for i in range(0, nparam):
cmd_frmt += f" --p{i} {'{}'}"
problem = Problem(cmd_frmt)
a, b = -15, 30
problem.spec_dim(p_id=0, p_space=(a, b), default=a)
problem.spec_dim(p_id=1, p_space=(a, b), default=a)
problem.spec_dim(p_id=2, p_space=[a+i for i in range(b-a)], default=a)
problem.spec_dim(p_id=3, p_space=[a+i for i in range(b-a)], default=a)
problem.spec_dim(p_id=4, p_space= list(np.random.permutation([str(a+i) for i in range(b-a)])), default=str(a))
problem.spec_dim(p_id=5, p_space= list(np.random.permutation([str(a+i) for i in range(b-a)])), default=str(a))
problem.checkcfg()
if __name__ == '__main__':
print(problem)
The executable.py file defines the method to evaluate a point in the search space:
#!/usr/bin/env python
from __future__ import print_function
import re
import os
import sys
import time
import json
import math
import os
import argparse
import numpy as np
from numpy import abs, cos, exp, mean, pi, prod, sin, sqrt, sum
seed = 12345
def create_parser():
'command line parser'
parser = argparse.ArgumentParser(add_help=True)
group = parser.add_argument_group('required arguments')
parser.add_argument('--p0', action='store', dest='p0',
nargs='?', const=2, type=float, default=-15.0,
help='parameter p0 value')
parser.add_argument('--p1', action='store', dest='p1',
nargs='?', const=2, type=float, default=-15.0,
help='parameter p1 value')
parser.add_argument('--p2', action='store', dest='p2',
nargs='?', const=2, type=int, default=-15,
help='parameter p2 value')
parser.add_argument('--p3', action='store', dest='p3',
nargs='?', const=2, type=int, default=-15,
help='parameter p3 value')
parser.add_argument('--p4', action='store', dest='p4',
nargs='?', const=2, type=str, default='-15',
help='parameter p4 value')
parser.add_argument('--p5', action='store', dest='p5',
nargs='?', const=2, type=str, default='-15',
help='parameter p5 value')
return(parser)
parser = create_parser()
cmdline_args = parser.parse_args()
param_dict = vars(cmdline_args)
print(param_dict)
p0 = param_dict['p0']
p1 = param_dict['p1']
p2 = param_dict['p2']
p3 = param_dict['p3']
p4 = int(param_dict['p4'])
p5 = int(param_dict['p5'])
x=np.array([p0,p1,p2,p3,p4,p5])
def ackley( x, a=20, b=0.2, c=2*pi ):
x = np.asarray_chkfinite(x) # ValueError if any NaN or Inf
n = len(x)
s1 = sum( x**2 )
s2 = sum(cos( c * x ))
return -a*exp( -b*sqrt( s1 / n )) - exp( s2 / n ) + a + exp(1)
pval = ackley(x, a=20, b=0.2, c=2*pi)
print('OUTPUT:%1.3f'%pval)
Reinforcement learning based search with proximal policy optimization
mpirun -np 2 python -m ytopt.search.ppo_a3c --prob_path=<PROBLEM_DIR_PATH>/problem.py --exp_dir=<EXP_DIR_PATH> --prob_attr=problem --exp_id=<ID> --max_time=60 --base_estimator='PPO'
Bayesian optimization with random forest model
mpirun -np 2 python -m ytopt.search.async_search --prob_path=<PROBLEM_DIR_PATH>/problem.py --exp_dir=<EXP_DIR_PATH> --prob_attr=problem --exp_id=<ID> --max_time=60 --base_estimator='RF'
Random search
mpirun -np 2 python -m ytopt.search.async_search --prob_path=<PROBLEM_DIR_PATH>/problem.py --exp_dir=<EXP_DIR_PATH> --prob_attr=problem --exp_id=<ID> --max_time=60 --base_estimator='DUMMY'
- GitHub repository: https://github.com/ytopt-team/ytopt
The core ytopt team is at Argonne National Laboratory:
- Prasanna Balaprakash [email protected], Lead and founder
- Romain Egele [email protected]
- Paul Hovland [email protected]
Modules, patches (code, documentation, etc.) contributed by:
Questions, comments, feature requests, bug reports, etc. can be directed to:
-
Our mailing list: [email protected] or https://groups.io/g/ytopt
-
Issues on GitHub
Patches are much appreciated on the software itself as well as documentation. Optionally, please include in your first patch a credit for yourself in the list above.
The ytopt team uses git-flow to organize the development: Git-Flow cheatsheet. For tests we are using: Pytest.
- YTune: Autotuning Compiler Technology for Cross-Architecture Transformation and Code Generation, U.S. Department of Energy Exascale Computing Project (2017--Present)
- Scalable Data-Efficient Learning for Scientific Domains, U.S. Department of Energy 2018 Early Career Award funded by the Advanced Scientific Computing Research program within the DOE Office of Science (2018--Present)
TBD
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent