Getting started

Install the package in your Python environment:

pip install skgrad

The runtime requires Python 3.9 or later, NumPy, SciPy, scikit-learn, and threadpoolctl. Poisson-loss MLP regression requires scikit-learn 1.7 or later. The documentation toolchain uses Python 3.12 in CI.

Your first input gradient

This binary classifier differentiates its decision score, not its probability. A positive derivative means increasing that input increases the score for model.classes_[1]. Each row corresponds to an observation, each column to an input feature.

import numpy as np
from sklearn.linear_model import LogisticRegression

import skgrad


X = np.array([[0.0, 0.0], [1.0, 0.0], [0.0, 1.0], [1.0, 1.0]])
y = np.array([0, 1, 0, 1])
model = LogisticRegression().fit(X, y)

scores, jacobian = skgrad.value_and_jacobian(model, [[0.25, 0.75]])
print(scores)
print(jacobian)

Run this file from an installed checkout with python examples/basic.py. The gradient of this affine score is the same at every input, although the score itself changes.

Choose the right result

Function	Result
model_output(model, X)	Values with shape (samples, outputs)
input_gradient(model, X, target=...)	One output’s gradient, (samples, features)
input_jacobian(model, X)	All output derivatives, (samples, outputs, features)
value_and_jacobian(model, X)	Named pair with .values and .jacobian

Omit target only for a single output. For multiclass and multioutput models, select an integer output position or request the full Jacobian.

supports(model) checks whether an analytic backend exists. It does not certify that the model is fitted. Fit the estimator first, then pass finite dense inputs in the same feature order and units used at fitting time.