Changelog • kerasnip

kerasnip (development version)

Fixed predict() failing with “Model not registered” after saving and reloading a kerasnip workflow in a new R session (#38). predict() now automatically replays the parsnip registration from metadata stored on the spec — no manual step required after bundle::unbundle() or readRDS().
Fixed get_keras_object() returning bare class constructors instead of instances for loss and metric objects, which caused save_model() to fail when those objects were passed to compile() (#42).
Fixed predict() and keras_evaluate() / extract_keras_model() silently failing when the Python external pointer became invalid after an RDS round-trip. Both functions now detect the invalid pointer via reticulate::py_validate_xptr() and transparently restore the model from the serialized bytes stored in the fit object (#42).
Fixed CRAN NOTE: added missing importFrom(stats, predict) so predict.kerasnip_model_fit is correctly resolved from the stats namespace.
Fixed compile_keras_grid() crashing with a vctrs_error_subscript_oob error when passed a zero-row tibble (e.g. tibble::tibble()). The function now stops early with an informative message. Use tibble::tibble(.rows = 1L) to build the model once with the spec’s current arguments and no hyperparameter variation.

Every spec instance now carries the kerasnip_spec class and embedded metadata (kerasnip_layer_blocks, kerasnip_functional), enabling transparent auto-registration on predict (closes #39).
fit() on a kerasnip spec now tags the result with kerasnip_model_fit class to enable the auto-registration dispatch.
At fit time the Keras model is serialized to a raw byte vector (.keras format) stored in the model_fit object. This makes plain saveRDS() / readRDS() fully supported without any extra steps (#42).
bundle::bundle() / unbundle() is now also supported as an alternative persistence strategy for MLOps and deployment workflows (#42).

Added the saving_and_reloading vignette explaining both the saveRDS and bundle workflows, with a comparison table and a description of the auto-restore mechanism (#42).
Corrected the “Save and Reload” sections in the Sequential Workflows and Functional Workflows vignettes, which previously stated that saveRDS does not work (#42).
Added a @section to both spec function reference pages explaining the bundle::bundle() workflow (closes #40).

CRAN release: 2025-12-06

inp_spec() now interprets named vectors in an argument-first orientation (c(input_a = "processed_1")). Existing code that used the previous upstream-first style must swap the names and values.

Updated README, vignettes, and reference docs to reflect the new inp_spec() mapping semantics and added guidance for migrating older code.

Added a regression test that fails fast when the legacy mapping orientation is supplied.

CRAN release: 2025-09-18

Added comprehensive end-to-end tests for tuning fit_* and compile_* parameters, and for autoplot uniqueness with multiple similar parameters.

Added new vignettes:
- “Transfer Learning with Keras Applications”
- “Tuning Multiple Similar Parameters: Ensuring autoplot Uniqueness”
- “Tuning Fit and Compile Arguments”

Enhanced register_model_args to improve matching of Keras arguments to dials functions and correctly assign package sources for dials parameters.
Refined remove_keras_spec to be more precise in removing model specifications, preventing unintended removal of other objects.

Fixed a bug in the documentation where examples were not self-contained, causing issues with CRAN checks. This involved updating examples to be fully runnable and cleaning up created model specifications (#22).
As part of this fix, a new helper function model_exists() was introduced and exported.

CRAN release: 2025-09-03

Added support for functional API (create_keras_functional_spec()) (#6).
Introduced custom steps, including step_collapse for collapsing columns in list columns (#20).
Added evaluation helpers: keras_evaluate(), extractors for summary and history (#12).
Introduced modularized helpers for build and compile of keras models (#14).
Added sequential workflow and functional workflow examples and vignettes (#20).
Added new tests to improve coverage.
Added pkgdown site improvements (favicon, documentation pages, guides) (#10, #16, #18 and #20).

Fixed issues with compile_ and fit_ argument handling.
Fixed issues with predict() and evaluate() to handle multiple outputs correctly (#18).
Fixed documentation typos and pkgdown errors.
Fixed utils issues and missing dependencies.
Fixed warnings and CRAN check issues.