martijnbosma / p3gomea-mis

P3GOMEA for medical image segmentation

SAGOMEA is an efficient surrogate-assisted Evolutionary Algorithm for combinatorial (discrete) optimization problems with expensive fitness functions. The paper (preprint): https://arxiv.org/abs/2104.08048

In this repository you can also find implementations of non-surrogate search algorithms, such as GOMEA, Local Search (LS), and Random Search (RS).

GOMEA for NAS for Medical Image Segmentation CUDA_VISIBLE_DEVICES=1 ./GOMEA --L=30 --timeLimit=17200000 --maxEvals=1500 --functionName=NAS --alphabet=py_src/../alphabets/NAS_MIS.txt --folder=py_src/../results/GOMEA_corr06_1/ --seed=0

Installation

1. pip install -r requirements.txt
2. Set Python versions in Makefile_SAGOMEA and Makefile_GOMEA to the Python paths of your system
3. To compile SAGOMEA: make -f Makefile_SAGOMEA To compile GOMEA: make -f Makefile_GOMEA

SAGOMEA

• Usage information is shown if ./SAGOMEA --help is typed
• An example of how to run SAGOMEA with the default hyperparameters is specified in the function run_SAGOMEA in the run_algorithms.py file
• You can specify a surrogate model type used by SAGOMEA:

1. Support Vector Regression (SVR)
2. Random Forest (RF)
3. Gradient Boosting (Catboost Regressor)
4. Multilayer Perceptron (MLP)

• All surrogate models are defined in the file: py_src/surrogateModel.py
• The recommended value of hyperparameter is 0.999

Using a custom fitness function

To use SAGOMEA for optimizing your own fitness function, it needs to be specified in py_src/fitnessFunctions.py

• Inherit a function class from the DummyFitnessFunction class
• If necessary, modify the constructor
• Specify fitness(self, x) function
• Note that a logger class instance should be used to save evaluated solutions (an example is shown in DummyFitnessFunction) .
• All obtained solutions during an optimization run along with their fitness values are stored in the file folder/optimization.txt

Running search algorithms in parallel

Parallel runs of SAGOMEA (or other search algorithms) can be done using run_algorithms.py script

• Specify problems and algorithms variables in run_algorithms.py. An example is provided in lines 329-330.
• For example, running python3 run_algorithms.py test 3600 5000 0 50 10 would execute 50 runs (with ids from 0 to 49) of the specified search algorithm(s) on the specified search problem(s), with 3600 seconds time limit (per run), 5000 fitness evaluations (per run), using test root folder, and performing 10 runs in parallel.