gentun

Python package for distributed genetic algorithm-based hyperparameter tuning

Table of Contents

1. About The Project
2. Installation
3. Usage
   • Single Node
     • Adding Pre-defined Individuals
     • Performing a Grid Search
   • Multiple Nodes
     • Redis Setup
     • Controller Node
     • Worker Nodes
4. Supported Models
5. Contributing
6. References

About The Project

The goal of this project is to create a simple framework for hyperparameter tuning of machine learning models, like Neural Networks and Gradient Boosting Trees, using a genetic algorithm. Evaluating the fitness of an individual in a population requires training a model with a specific set of hyperparameters, which is a time-consuming task. To address this issue, we offer a controller-worker system: multiple workers can perform model training and cross-validation of individuals provided by a controller while this controller manages the generation of offspring through reproduction and mutation.

"Parameter tuning is a dark art in machine learning, the optimal parameters of a model can depend on many scenarios." ~ XGBoost tutorial on Parameter Tuning

"The number of possible network structures increases exponentially with the number of layers in the network, which inspires us to adopt the genetic algorithm to efficiently traverse this large search space." ~ Genetic CNN paper

Installation

pip install gentun

Some model handlers require additional libraries. You can also install their dependencies with:

pip install "gentun[xgboost]"  # or "gentun[tensorflow]"

To setup a development environment, run:

python -m pip install --upgrade pip
pip install 'flit>=3.8.0'
flit install --deps develop --extras tensorflow,xgboost

Usage

Single Node

The most basic way to run the algorithm is using a single machine, as shown in the following example where we use it to find the optimal hyperparameters of an xgboost model. First, we download a sample dataset:

from sklearn.datasets import load_iris

data = load_iris()
x_train = data.data
y_train = data.target

Next, we need to define the hyperparameters we want to optimize:

from gentun.genes import RandomChoice, RandomLogUniform

genes = [
    RandomLogUniform("learning_rate", minimum=0.001, maximum=0.1, base=10),
    RandomChoice("max_depth", [3, 4, 5, 6, 7, 8, 9, 10]),
    RandomChoice("min_child_weight", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]),
]

We are using the gentun.models.xgboost.XGBoost handler, which performs k-fold cross validation with available train data and returns an average metric over the folds. Thus, we need to define some static parameters which are shared across the population over all generations:

kwargs = {
    "booster": "gbtree",
    "objective": "multi:softmax",
    "metrics": "mlogloss",  # The metric we want to minimize with the algorithm
    "num_class": 3,
    "nfold": 5,
    "num_boost_round": 5000,
    "early_stopping_rounds": 100,
}

Finally, we are ready to run our genetic algorithm. gentun will check that all the model's required parameters are passed either through genes or keyword arguments.

from gentun.algorithms import Tournament
from gentun.models.xgboost import XGBoost
from gentun.populations import Population

# Run the genetic algorithm with a population of 50 for 100 generations
population = Population(genes, XGBoost, 50, x_train, y_train, **kwargs)
algorithm = Tournament(population)
algorithm.run(100, maximize=False)

As shown above, when the model and genes are implemented, experimenting with the genetic algorithm is simple. See for example how easily can the Genetic CNN paper be defined on the MNIST handwritten digits set.

Note that in genetic algorithms, the fitness of an individual is a number to be maximized. By default, this framework follows this convention. Nonetheless, to make the framework more flexible, you can use the maximize=False parameter in algorithm.run() to override this behavior and minimize your fitness metric (e.g. when you want to minimize the loss, for example rmse or binary crossentropy).

Adding Pre-defined Individuals

Oftentimes, it's convenient to initialize the genetic algorithm with some known individuals instead of a random population. You can add custom individuals to the population before running the genetic algorithm if you already have an intuition of which hyperparameters work well with your model:

from gentun.models.xgboost import XGBoost
from gentun.populations import Population


# Best known parameters
hyperparams = {
    "learning_rate": 0.1,
    "max_depth": 9,
    "min_child_weight": 1,
}

# Generate a random population and then add a custom individual
population = Population(genes, XGBoost, 49, x_train, y_train, **kwargs)
population.add_individual(hyperparams)

Performing a Grid Search

Grid search is also widely used for hyperparameter optimization. This framework provides gentun.populations.Grid, which can be used to conduct a grid search over a single generation pass. You must use genes which define the sample() method, so that uniformly distributed hyperparameter values are obtained with it.

from gentun.genes import RandomChoice, RandomLogUniform
from gentun.models.xgboost import XGBoost
from gentun.populations import Grid


genes = [
    RandomLogUniform("learning_rate", minimum=0.001, maximum=0.1, base=10),
    RandomChoice("max_depth", [3, 4, 5, 6, 7, 8, 9, 10]),
    RandomChoice("min_child_weight", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]),
]

gene_samples = [10, 8, 11]  # How many samples we want to get from each gene

# Generate a grid of individuals
population = Grid(genes, XGBoost, gene_samples, x_train, y_train, **kwargs)

Running the genetic algorithm on this population for just one generation is equivalent to doing a grid search over 10 learning_rate values, all max_depth values between 3 and 10, and all min_child_weight values between 0 and 10.

Multiple Nodes

You can speed up the genetic algorithm by using several machines to evaluate individuals in parallel. One of node has to act as a controller, generating populations and running the genetic algorithm. Each time this controller node needs to evaluate an individual from a population, it will send a request to a job queue that is processed by workers which receive the model's hyperparameters and perform model fitting through k-fold cross-validation. The more workers you run, the faster the algorithm will evolve each generation.

Redis Setup

The simplest way to start the Redis service that will host the communication queues is through docker:

docker run -d --rm --name gentun-redis -p 6379:6379 redis

Controller Node

To run the distributed genetic algorithm, define a gentun.services.RedisController and pass it to the Population instead of the x_train and y_train data. When the algorithm needs to evaluate the fittest individual, it will pass the hyperparameters to a job queue in Redis and wait till all the individual's fitness are evaluated by worker processes. Once this is done, the mutation and reproduction steps are run by the controller and a new generation is produced.

from gentun.models.xgboost import XGBoost
from gentun.services import RedisController

controller = RedisController("experiment", host="localhost", port=6379)
# ... define genes
population = Population(genes, XGBoost, 100, controller=controller, **kwargs)
# ... run algorithm

Worker Nodes

The worker nodes are defined using the gentun.services.RedisWorker class and passing the handler to it. Then, we use its run() method with train data to begin processing jobs from the queue. You can use as many nodes as desired as long as they have network access to the redis server.

from gentun.models.xgboost import XGBoost
from gentun.services import RedisWorker

worker = RedisWorker("experiment", XGBoost, host="localhost", port=6379)

# ... fetch x_train and y_train
worker.run(x_train, y_train)

Supported Models

This project supports hyperparameter tuning for the following models:

• XGBoost regressor and classifier
• Scikit-learn regressor and classifier
• Genetic CNN with Tensorflow
• A Genetic Programming Approach to Designing Convolutional Neural Network Architectures

Contributing

We welcome contributions to enhance this library. You can submit your custom subclasses for:

• gentun.models.Handler
• gentun.genes.Gene

Our roadmap includes:

• Training data sharing between the controller and worker nodes
• Proof-of-work validation of what worker nodes submit

You can also help us speed up hyperparameter search by contributing your spare GPU time.

For more details on how to contribute, please check our contribution guidelines.

References

Genetic Algorithms

• Artificial Intelligence: A Modern Approach. 3rd edition. Section 4.1.4
• https://github.com/DEAP/deap
• http://www.theprojectspot.com/tutorial-post/creating-a-genetic-algorithm-for-beginners/3

XGBoost Parameter Tuning

• http://xgboost.readthedocs.io/en/latest/parameter.html
• http://xgboost.readthedocs.io/en/latest/how_to/param_tuning.html
• https://www.analyticsvidhya.com/blog/2016/03/complete-guide-parameter-tuning-xgboost-with-codes-python/

Papers

• Lingxi Xie and Alan L. Yuille, Genetic CNN
• Masanori Suganuma, Shinichi Shirakawa, and Tomoharu Nagao, A Genetic Programming Approach to Designing Convolutional Neural Network Architectures