get_default_training_options: Get default training options

View source: R/prepare_mofa.R

get_default_training_optionsR Documentation

Get default training options


Function to obtain the default training options.





an untrained MOFA


This function provides a default set of training options that can be modified and passed to the MOFA object in the prepare_mofa step (see example), i.e. after creating a MOFA object (using create_mofa) and before starting the training (using run_mofa) The training options are the following:

  • maxiter: numeric value indicating the maximum number of iterations. Default is 1000. Convergence is assessed using the ELBO statistic.

  • drop_factor_threshold: numeric indicating the threshold on fraction of variance explained to consider a factor inactive and drop it from the model. For example, a value of 0.01 implies that factors explaining less than 1% of variance (in each view) will be dropped. Default is -1 (no dropping of factors)

  • convergence_mode: character indicating the convergence criteria, either "fast", "medium" or "slow", corresponding to 0.0005%, 0.00005% or 0.000005% deltaELBO change.

  • verbose: logical indicating whether to generate a verbose output.

  • startELBO: integer indicating the first iteration to compute the ELBO (default is 1).

  • freqELBO: integer indicating the first iteration to compute the ELBO (default is 1).

  • stochastic: logical indicating whether to use stochastic variational inference (only required for very big data sets, default is FALSE).

  • gpu_mode: logical indicating whether to use GPUs (see details).

  • seed: numeric indicating the seed for reproducibility (default is 42).


Returns a list with default training options


# Using an existing simulated data with two groups and two views
file <- system.file("extdata", "test_data.RData", package = "MOFA2")

# Load data dt (in data.frame format)

# Create the MOFA object
MOFAmodel <- create_mofa(dt)

# Load default training options
train_opts <- get_default_training_options(MOFAmodel)

# Edit some of the training options
train_opts$convergence_mode <- "medium"
train_opts$startELBO <- 100
train_opts$seed <- 42

# Prepare the MOFA object
MOFAmodel <- prepare_mofa(MOFAmodel, training_options = train_opts)

bioFAM/MOFA2 documentation built on March 21, 2023, 5:27 p.m.