Skip to contents

Extract linear-layer weights and map them to feature / class names

get_weights.Rd

Reads a model exported by the Rust Burn pipeline together with its companion metadata file and returns a tidy weight matrix whose rows correspond to the original feature names and whose columns correspond to the class labels.

Usage

get_weights(dir)

Arguments

dir

character(1) Path to the artefact directory that contains both files produced by run_custom():

  • model.mpk – tensor weights saved via NamedMpkFileRecorder::<FullPrecisionSettings>

  • meta.mpk – MessagePack blob created by save_artifacts() holding feature_names and class_labels.

Value

A base-data.frame with dimension \((\#\;features) \times (\#\;classes)\), where rownames(wmat) are the feature names and colnames(wmat) are the class labels. Cell (i,j) is the weight connecting feature i to logit j.

Details

Internally the function:

  1. deserialises the two MessagePack files with msgpackR;

  2. raw‐decodes the tensor bytes through decode_param();

  3. reshapes the flat vector into a column-major matrix using the stored shape ([out_dim, in_dim]);

  4. transposes it so that rows align with features;

  5. re-labels rows and columns from the metadata lists.

The resulting object is ready for pheatmap(), corrplot(), or as.matrix() for further analysis.

See also

  • msgpack_read() from msgpackR – generic MessagePack reader

  • decode_param() – helper that converts Burn tensor blobs into R vectors

Examples

if (FALSE) { # \dontrun{
w <- get_weights("artifacts/run-42")
head(w[, 1:5])         # first 5 classes

# visualise top positive / negative features for class 3
cls <- 3
w_sorted <- w[order(w[, cls]), cls]
barplot(tail(w_sorted, 10), horiz = TRUE, las = 1)
barplot(head(w_sorted, 10), horiz = TRUE, las = 1)
} # }