Skops

0.13.0 · active · verified Thu Apr 09

Skops is a set of tools designed for machine learning in production, focusing on model persistence, sharing, and documentation. It provides secure object dumping/loading, integrates with Hugging Face Hub for sharing, and helps generate Model Cards. The library is currently at version 0.13.0 and maintains a regular release cadence, typically monthly or bi-monthly, addressing bug fixes, enhancements, and security improvements.

Warnings

breaking Loading untrusted models or Model Cards from external sources can lead to arbitrary code execution due to pickle's security vulnerabilities. Skops offers security mechanisms like the `trusted` argument for `skops.io.load` and `allow_insecure_loading` for `skops.card.Card`.
Fix: Always set `trusted=False` (which is the default) for `skops.io.load` when loading from untrusted sources. For `skops.card.Card`, explicitly set `allow_insecure_loading=False` (available from v0.13.0) or ensure the source is trusted. Carefully review the source of any model or card before loading.
gotcha Model persistence is highly sensitive to library versions. Models saved with one version of `skops`, `scikit-learn`, `numpy`, or other dependencies might not load or behave identically with significantly different versions.
Fix: Document the exact versions of `skops` and all major dependencies (`scikit-learn`, `numpy`, `pandas`, etc.) used to save a model. Ideally, recreate the environment using tools like `pip-tools` or `conda` to ensure consistent dependency versions when loading.
gotcha The `skops convert` CLI command, designed to convert model files (e.g., `joblib` to `skops`), is a 'best-effort' attempt and does not guarantee an exact, bit-for-bit identical conversion.
Fix: After using `skops convert`, always thoroughly test and verify the converted model's behavior and predictions to ensure it functions as expected. Consider the original file the definitive source unless verified.

Install

pip install skops Install latest version

Imports

dump
```
from skops.io import dump
```
load
```
from skops.io import load
```
Card
```
from skops.card import Card
```
push
```
from skops.hub import push
```

Quickstart

This quickstart demonstrates how to train a basic scikit-learn model, save it using `skops.io.dump`, and then load it back using `skops.io.load`. This is the most common use case for skops, providing a secure and robust way to persist machine learning models.

from sklearn.linear_model import LogisticRegression
from skops.io import dump, load

# Train a simple model
model = LogisticRegression(random_state=42)
model.fit([[0, 0], [0, 1], [1, 0], [1, 1]], [0, 0, 1, 1])

# Save the model securely
dump(model, 'model.skops')

# Load the model
loaded_model = load('model.skops')

# Make a prediction with the loaded model
prediction = loaded_model.predict([[0, 0]])
print(f"Original model class: {type(model).__name__}")
print(f"Loaded model class: {type(loaded_model).__name__}")
print(f"Prediction: {prediction[0]}")

view raw JSON →