Reads a .segmantR archive created by sg_package_model() and
returns an sg_trained_model object.
Details
Before extracting, entry names are checked (no absolute paths, parent
segments, backslashes or drive letters; entry count and total size
limits). After extraction, symbolic links are rejected, checksums.sha256
is verified (missing, extra or modified files are errors), the model card
schema and major version and the backend are checked. Archives written by
segmantR 0.1.0 have no checksums; they load with integrity = "unverified" unless allow_legacy = FALSE.
Examples
# a dedicated directory with a stand-in weight file
weights <- file.path(tempfile("model"), "cellpose_weights")
dir.create(dirname(weights))
writeBin(as.raw(1:16), weights)
mdl <- new_sg_trained_model(weights, backend = "cellpose",
base_model = "cyto3",
training_metrics = list(n_epochs = 50L))
archive <- sg_package_model(mdl, tempfile(fileext = ".segmantR"))
#> ✔ Model packaged to /tmp/RtmpZNYdOa/file1f6c50426c71.segmantR.
model <- sg_load_model(archive)
#> ✔ Loaded model from /tmp/RtmpZNYdOa/file1f6c50426c71.segmantR.
#> ℹ Backend: cellpose, base: cyto3.
model$integrity
#> [1] "verified"