For hyper-parameter tuning, this project implements a flexible, distributed and parallel framework for robust Bayesian Optimization, and the relative module Tuner is based on the project RoBo from AutoML.
import numpy as np
import logging
logging.basicConfig(level=logging.INFO)
from tuner.tuner import Tuner
from test_model import TestModel
# Defining the bounds and dimensions of the input space
lower = np.array([0, 2, 1])
upper = np.array([6, 5, 9])
# Start Bayesian optimization to optimize the objective function
tuners = Tuner(TestModel.train, lower, upper, num_iter=10, num_worker=4)
results = tuners.run()AutoTF is based on project RoBo, and uses the Gaussian processes library george and the random forests library pyrfr. In order to use this library make sure the libeigen and swig are installed:
sudo apt-get install libeigen3-dev swig gfortran
Before you install AutoTF you have to install the required dependencies. We use a for loop because we want to preserve the installation order of the list of dependencies in the requirments file.
for req in $(cat requirements.txt); do pip install $req; done
This will install the basis requirements that you need to run AutoTF’s core functionality. If you want to make use of the full functionality (for instance Bohamiann, Fabolas, …) you can install all necessary dependencies by:
for req in $(cat all_requirements.txt); do pip install $req; done
Note: This may take a while to install all dependencies.