Approximately Exact Line Search
utils.py contains the implementation of Approximately Exact Line Search, as well as various baseline methods and helper functions.
profile.py runs the line search methods on the benchmark problems, and saves the results. Options are:
--problems_file: the .txt file describing the desired problems to run
--objective: either dfo or logistic
--recompute: pass this option to force recomputation (prior results are reused by default)
--threads: number of processes to run concurrently; for accurate timing data, we recommend fewer processes than cores
--verbose: pass this option to print progress details (recommend piping to a file)
make_profiles.py generates the figures in our paper. We recommend running profile.py before make_profiles.py. Options are:
--type: dfo (benchmark DFO problems), batchsize (logistic regression with various minibatch sizes), or epsilon (logistic regression with various required accuracies)
--threads: number of processes to run concurrently
check.py can be run concurrently with profile.py to monitor progress. check.py prints which experiments are in progress or remain to be initiated. Options are:
--problems_file: the .txt file describing the desired problems to run
--objective: either dfo or logistic
--clean: pass this option to delete logfiles for any unfinished experiments
--quiet: pass this option to print only logfile names for unfinished experiments
problems/ contains .txt files describing the benchmark problems.
data/ contains logistic regression benchmark data from the UCI Adult dataset (https://archive.ics.uci.edu/ml/datasets/Adult) and the KDD-Cup 2004 Physics and Biology datasets (https://www.kdd.org/kdd-cup/view/kdd-cup-2004/Tasks).
calfun.py, dfovec.py, and dfoxs.py contain Python implementations of the DFO benchmark problems at https://www.mcs.anl.gov/~more/dfo/.
To run the DFO experiments and generate logfiles:
python profile.py --problems_file=problems/dfo.txt --objective=dfo
While that is running, in another window you can monitor progress by running:
python check.py --problems_file=problems/dfo.txt --objective=dfo
To plot the performance profile:
python make_profiles.py --type=dfo
This repository contains the code to reproduce results in the following paper:
https://arxiv.org/abs/2011.04721