Migration from personalized-srm¶

Use a separate environment for the new package and retain the frozen research environment for historical replay. This package is multimodalsrm, not a new release of the personalized_srm namespace.

Imports¶

Update imports and install the Bayesian extra when needed. Data nesting, response coordinates, estimator parameters and target-exclusion semantics are retained by the extraction. Legacy estimator catalogs and research experiment runners are not migrated. No alias package for the old namespace is provided.

Fitted GP model archives¶

Use save_model and load_model from multimodalsrm.bayesian.workflow. The JSON/numeric archive format and historical serialized identifiers intentionally retain their personalized-srm names. Do not rewrite those strings to match the Python namespace.

Schemas 1 and 2 represent MAP group and MAP participant states; schema 3 represents training posteriors; schema 4 represents updated posterior states. The loader reconstructs a fitted model for prediction/reporting without a new parameter optimization or MCMC run. Saved diagnostics retain their original success, failure or undefined status. The extraction check replayed source-generated synthetic archives for all four schemas under the new namespace, including calibrated predictions and seeded posterior paths. See the validation record. Successful replay does not requalify the old scientific result under a new runtime.

Keep original archives unchanged. Archives can contain native observations and posterior draws, so store them with the research data rather than committing them to the package. Saving requires a new destination. Preserve any training standardizer with save_model(..., standardizer=scaler) and use the restored standardizer consistently. The MAP quickstart performs a synthetic save/load comparison.

SearchConfig(r_init=True) is a new opt-in option that uses a bounded R fit to initialize the first MAP restart. It is not the default: new searches keep the historical data-based first start, so an existing workflow is unaffected. Archives created before this option retain r_init=False when loaded, including models with an implicit search configuration, and their saved fit evidence is preserved.

R pickle/joblib models¶

Python pickle/joblib stores class module paths, including paths under personalized_srm.multimodal. Installing multimodalsrm alone does not resolve those paths. There is no automatic converter or compatibility unpickler in this package.

For an existing trusted R model, retain the original source checkout, dependency versions and preprocessing artifacts. Load it and make predictions in that original environment. Export only the numeric outputs and metadata needed for downstream analysis. For new-package use, refit from the authorized original training data with the same configuration, then compare objective, kernels, validity masks and predictions against the old environment under explicit tolerances. A refit is a new fitted artifact; preserve its provenance and do not call it a converted old model. If original data are unavailable, continue prediction/replay in the old environment.

Warmup checkpoints¶

Warmup checkpoints are different from fitted GP archives. They restore the complete sampler state at the end of warmup and restart retained sampling from that boundary. They do not append partially retained draws. Repeated/resumed trajectories must not be combined as independent samples.

Compatibility checks hash every Python source file under the model package tree, as well as recording numerical runtime, devices, PRNG configuration and target identity. The namespace extraction changes that source tree even when mathematical behavior is preserved. An existing research-repository warmup checkpoint must resume in its original frozen source and runtime environment. Do not edit checkpoint identities, remove hashes or assume fitted-archive compatibility implies warmup compatibility.

New package checkpoints can be created with fit(..., warmup_checkpoint=path) and resumed with the identical configuration, data, explicit integer seed, runtime and source using resume_warmup=True. This API covers full dense/grouped posterior training with independent noise and no run baselines; it does not cover posterior updates, conditional parameter targets or state-space inference. Checkpoints require new destinations and never overwrite existing ones. Retain the source/environment for each long run.

Skin-conductance responses¶

Release 0.2.0 standardizes SCR workflows on BatemanSCR and removes BachSCR. This is a breaking change to imports, fitting and Bach archive replay. GP archives containing Bach records fail with migration guidance; they are not converted to Bateman. The archive schema and remaining kernel identifiers stay unchanged.

Reproduce old Bach analyses in their original environment. Release 0.1.0 supports Bach, and development commit 8eeb84fcc8d2c6a37f36ce748a13e6c3782e0ce8 is the snapshot immediately before removal. This also applies to R pickle/joblib models and warmup checkpoints. For current workflows, configure Bateman response bounds and GP priors and refit from original observations. Shape parameters, loadings and lag estimates are not automatically transferable. See the SCR guide.