Performs hyperparameter optimization for CISS-VAE using Optuna, with support for Rich progress bars and flexible layer arrangement strategies.
Usage
autotune_cissvae(
data,
index_col = NULL,
val_proportion = 0.1,
replacement_value = 0,
columns_ignore = NULL,
clusters,
save_model_path = NULL,
save_search_space_path = NULL,
n_trials = 20,
study_name = "vae_autotune",
device_preference = "cuda",
show_progress = FALSE,
optuna_dashboard_db = NULL,
load_if_exists = TRUE,
seed = 42,
verbose = FALSE,
constant_layer_size = FALSE,
evaluate_all_orders = FALSE,
max_exhaustive_orders = 100,
num_hidden_layers = c(1, 4),
hidden_dims = c(64, 512),
latent_dim = c(10, 100),
latent_shared = c(TRUE, FALSE),
output_shared = c(TRUE, FALSE),
lr = c(1e-04, 0.001),
decay_factor = c(0.9, 0.999),
beta = 0.01,
num_epochs = 500,
batch_size = 4000,
num_shared_encode = c(0, 1, 3),
num_shared_decode = c(0, 1, 3),
encoder_shared_placement = c("at_end", "at_start", "alternating", "random"),
decoder_shared_placement = c("at_start", "at_end", "alternating", "random"),
refit_patience = 2,
refit_loops = 100,
epochs_per_loop = 500,
reset_lr_refit = c(TRUE, FALSE)
)Arguments
- data
Data frame or matrix containing the input data
- index_col
String name of index column to preserve (optional)
- clusters
Integer vector specifying cluster assignments for each row
- save_model_path
Optional path to save the best model's state_dict
- save_search_space_path
Optional path to save search space configuration
- n_trials
Number of Optuna trials to run
- study_name
Name identifier for the Optuna study
- device_preference
Preferred device ("cuda" or "cpu")
- show_progress
Whether to display Rich progress bars during training
- optuna_dashboard_db
RDB storage URL/file for Optuna dashboard
- load_if_exists
Whether to load existing study from storage
- seed
Base random seed for reproducible results
- verbose
Whether to print detailed diagnostic information
- constant_layer_size
Whether all hidden layers use same dimension
- evaluate_all_orders
Whether to test all possible layer arrangements
- max_exhaustive_orders
Max arrangements to test when evaluate_all_orders=TRUE
Numeric(2) vector: (min, max) for number of hidden layers
Numeric vector: hidden layer dimensions to test
- latent_dim
Numeric(2) vector: (min, max) for latent dimension
Logical vector: whether latent space is shared across clusters
Logical vector: whether output layer is shared across clusters
- lr
Numeric(2) vector: (min, max) learning rate range
- decay_factor
Numeric(2) vector: (min, max) LR decay factor range
- beta
Numeric: KL divergence weight (fixed or range)
- num_epochs
Integer: number of initial training epochs (fixed or range)
- batch_size
Integer: mini-batch size (fixed or range)
Numeric vector: numbers of shared encoder layers to test
Numeric vector: numbers of shared decoder layers to test
Character vector: placement strategies for encoder shared layers
Character vector: placement strategies for decoder shared layers
- refit_patience
Integer: early stopping patience for refit loops
- refit_loops
Integer: maximum number of refit loops
- epochs_per_loop
Integer: epochs per refit loop
- reset_lr_refit
Logical vector: whether to reset LR before refit