Saving and Using Saved Models

2026-04-08

This vignette shows how to save a trained CISS-VAE model, reload it in a later R session, and use it to impute new data.

As of rCISSVAE version 0.0.6, the recommended workflow is to save the model as a PyTorch state_dict together with an automatically generated architecture config file. This is more portable across Python and PyTorch and more robust than saving the full live Python object.

Saving a trained CISS-VAE model

library(reticulate)
library(rCISSVAE)

# Train a model
res <- run_cissvae(data, return_model = TRUE)

# Save the trained model to disk
# by default, this writes two files: 
# -  trained_vae.pt (the model weights)
# - trained_vae.pt.config.rds (the saved model configuration/architecture needed to rebuild the model when loading. )
save_cissvae_model(res$model, "trained_vae.pt", method="state_dict")

# IMPORTANT
# The Python environment must be active so 'torch' can be imported.

There is still the option to save the full model as a .pt file if desired by setting method='full'.

Loading a saved model and imputing data

library(rCISSVAE)
library(reticulate)

## Activate your Python environment 
reticulate::use_virtualenv("cissvae_environment", required = TRUE)

## Load full model object
model <- load_cissvae_model(
  file = "trained_vae.pt",
  method = "state_dict", ## (or 'full' if full was used when saving)
  device = "cpu"
)

## Perform imputation on new data
## Make sure your `data` has valid NAs and `clusters` vector is ready
## `val_proportion`, `categorical_column_map` and `replacement_value` are not needed because we are just imputing
imputed_df <- impute_with_cissvae(
  model_py = model,
  data = data,
  index_col = "index",
  columns_ignore = NULL,
  clusters = clusters,
  imputable_matrix = NULL,
  binary_feature_mask = NULL,
  batch_size = 4000L,
  seed = 42
)

# `imputed_df` is returned to R as a data.frame

If you have binary variables in your dataset, make sure to define the binary_feature_mask in your impute_with_cissvae() call and convert the probabilities for the binary variables into {0, 1} values after imputation using desired thresholding.