Skip to content

Saving and loading models

comprisk offers two ways to persist a fitted estimator. Both are lossless: round-tripped predictions are bit-identical to the in-memory model, enforced by the test suite.

model = CompetingRiskForest(n_estimators=500, random_state=42).fit(X, time, event)
model.save("model.crm")

# on the receiving side
import comprisk
model = comprisk.load("model.crm")

The file is a zip container of JSON metadata plus raw .npy arrays. Two properties make it the right format for sending models to collaborators:

  • No code execution on load. Unlike pickle, comprisk.load never unpickles objects — it reads JSON and arrays (np.load(allow_pickle=False)). A model file from outside your trust boundary cannot run code on your machine.
  • Deterministic bytes. Saving the same model twice in the same environment produces identical files (fixed zip timestamps), so a checksum recorded in a manifest stays valid across re-saves. Compressed bytes can differ across zlib builds, so verify checksums where the file was written.

FineGrayRegression supports the same save() / comprisk.load() pair. save() covers the default flat-tree forest; reference-mode and rfsrc-aligned forests (research and cross-library-parity configurations) persist via pickle instead.

Pickle / joblib

Standard pickling works as always and remains the right choice for sklearn-ecosystem plumbing (joblib caching, GridSearchCV, model registries that expect pickles). Old pickles from earlier comprisk versions load unchanged.

Sizes

Serialized forests are dominated by per-leaf competing-risk curves. comprisk stores these sparsely and rebuilds the dense representation on load (bit-identically), so files are far smaller than the in-memory footprint. Measured on a 20,000-sample, 20-tree synthetic fit (default parameters):

Serialization Size vs old format
Plain pickle, comprisk ≤ 0.7 (dense state) 202.8 MB 1×
Plain pickle, compact state 8.4 MB 24×
Compact state + joblib.dump(..., compress=("zlib", 3)) 4.8 MB 42×
save() (compact + deflate) 4.6 MB 44×

The ratio grows with forest size. On a real clinical forest (500 trees, min_samples_leaf=3, cohort size in the tens of thousands) whose dense-state pickle is 33.3 GB, measured 2026-08-17: the compact pickle is 444 MB (75×) and save() lands at 116 MB (287×) — written in 21 s, loaded in 40 s, with bit-identical round-tripped predictions and unchanged held-out concordance.

Loading rebuilds each tree's leaf CIF table from its sparse counts; this is a vectorized pass that adds seconds, not minutes, even for forests whose dense state runs to gigabytes.