sits_tuning: Tuning machine learning models hyper-parameters

In sits: Satellite Image Time Series Analysis for Earth Observation Data Cubes

Tuning machine learning models hyper-parameters

Description

Machine learning models use stochastic gradient descent (SGD) techniques to find optimal solutions. To perform SGD, models use optimization algorithms which have hyperparameters that have to be adjusted to achieve best performance for each application.

This function performs a random search on values of selected hyperparameters. Instead of performing an exhaustive test of all parameter combinations, it selecting them randomly. Validation is done using an independent set of samples or by a validation split. The function returns the best hyper-parameters in a list. Hyper-parameters passed to params parameter should be passed by calling sits_tuning_hparams().

Usage

sits_tuning(
  samples,
  samples_validation = NULL,
  validation_split = 0.2,
  ml_method = sits_tempcnn(),
  params = sits_tuning_hparams(optimizer = torchopt::optim_adamw, opt_hparams = list(lr =
    loguniform(10^-2, 10^-4))),
  trials = 30,
  multicores = 2,
  progress = FALSE
)

Arguments

samples	Time series set to be validated.
samples_validation	Time series set used for validation.
validation_split	Percent of original time series set to be used for validation (if samples_validation is NULL)
ml_method	Machine learning method.
params	List with hyper parameters to be passed to ml_method. User can use uniform, choice, randint, normal, lognormal, loguniform, and beta distribution functions to randomize parameters.
trials	Number of random trials to perform the random search.
multicores	Number of cores to process in parallel
progress	Show progress bar?

Value

A tibble containing all parameters used to train on each trial ordered by accuracy

Author(s)

Rolf Simoes, [e-mail]

References

James Bergstra, Yoshua Bengio, "Random Search for Hyper-Parameter Optimization". Journal of Machine Learning Research. 13: 281–305, 2012.

Examples

if (sits_run_examples()) {
    # find best learning rate parameters for TempCNN
    tuned <- sits_tuning(
        samples_modis_ndvi,
        ml_method = sits_tempcnn(),
        params = sits_tuning_hparams(
            optimizer = choice(
                torch::optim_adamw
            ),
            opt_hparams = list(
                lr = loguniform(10^-2, 10^-4)
            )
        ),
        trials = 4,
        multicores = 2,
        progress = FALSE
    )
    # obtain best accuracy, kappa and best_lr
    accuracy <- tuned$accuracy[[1]]
    kappa <- tuned$kappa[[1]]
    best_lr <- tuned$opt_hparams[[1]]$lr
}

sits documentation built on Nov. 2, 2023, 5:59 p.m.