Hyperopt Basics

This is an introduction to using Hyperopt library. I will use mostly terminology of machine learning (ML) as this library appeared in the ML community.

Hyperopt library is used to choose the hyperparameters, that is, parameters that must be set before the learning process. The learning process is the process of fitting a given model to some dataset, which is done by minimization of some function.

For example, when you fit model \(\hat f (x)\) by optimizing function \[ \frac{1}{N} \sum_{i=1}^N \left( y_i - \hat f (x_i) \right)^2 + \lambda R(f), \]

then \(\lambda\) is a hyperparameter that must be chosen before the optimization.

Overall algorithm of using Hyperopt

Optimization of the hyperparameters in Hyperopt consists of three steps:

1. Define an objective function that numerically measures the quality of the given set of hyperparameters.
2. Define hyperparameters space.
3. Call function hyperopt.fmin that will find optimal set of hyperparameters given objective function and hyperparameter space.

Defining objective function

The simplest way to define an objective function is the following:

def objective(params):
    # expand params, for example:
    alpha, beta = params
    # evaluate some nonnegative function `loss` using params
    loss = alpha**2 + beta**2

    return loss

Defining parameter space

Hyperparameter space is defined using utility module hp. For each parameter, we can specify which distribution it has.

For example, the following code specifies that variable \(\alpha \sim \uniform(-2, 5)\) while \(\beta \sim \normal(0, 3^2)\):

params_space = [
  hyperopt.hp.uniform('alpha', -2, 5),
  hyperopt.hp.normal('beta', 0, 3),
]

Running hyperparameter optimization

Now we are ready to optimize hyperparameters. For that, we use function hyperopt.fmin which accepts objective function and parameter space. Besides that, we need to specify the algorithm that we use, and how many trials we would like to run:

    params_star = hyperopt.fmin(
        objective,
        params_space,
        algo=hyperopt.tpe.suggest
    )

That’s it! This was an introduction to the optimization of hyperparameters using Hyperopt library.

This tutorial is based on the paper: https://conference.scipy.org/proceedings/scipy2013/pdfs/bergstra_hyperopt.pdf

• Howto
• Hyperopt