NCET

NCET (Neural-network Constraint Embedding Toolkit) converts supported PyTorch neural networks into exact mixed-integer linear constraints. Its graph-aware frontend preserves branches, shared tensors, and residual/skip connections instead of restricting models to sequential structures.

Installation

Install the published package from PyPI:

pip install ncet

For a local editable installation, run from the repository root:

pip install -e .

Minimal use

import cvxpy as cp
import numpy as np

from ncet import Bounds, form_milp


model.eval()
bounds = Bounds(
    lower=np.array([-1.0, -1.0]),
    upper=np.array([1.0, 1.0]),
)
encoding = form_milp(model, bounds, relu_binary_mode="reduced")

y = encoding.outputs[0]
problem = cp.Problem(cp.Maximize(y[0]), encoding.constraints)
problem.solve(solver=cp.SCIPY)

Input bounds do not include a batch dimension

Bounds describe exactly one sample. Use (features,) for an MLP or (channels, height, width) for an image, not (batch, features) or (batch, channels, height, width).

See the form_milp() user interface for every argument, accepted bound form, return field, and public exception.

Documentation map

For regular users, please refer to User interface, Examples, and Supported operators for detailed usage and examples.

For developers, please refer to Developer reference for detailed implementation details.

Mathematical background is available in Knowledge.

Section	Purpose
User interface	Build and consume an exact NCET encoding
Examples	Representative notebooks and applications
Supported operators	Authoritative current capability boundary
Exactness contract	Model assumptions and semantic guarantee
Mathematical background	Bound propagation and exact formulations
Developer reference	FX, GraphIR, propagation, and backend internals