globoptiml 0.3.0

Train neural networks via integer optimisation (CP-SAT) instead of gradient descent.

OptiML

Train neural networks via global mathematical optimisation (MINLP) instead of gradient descent.

OptiML translates a neural network architecture into a system of constraints and decision variables, then uses a MINLP solver (e.g. Couenne, SCIP) to find the weights that provably minimise the training loss. After solving, the model can be exported to PyTorch for inference.

Installation

pip install -e .            # core (numpy + pyomo)
pip install -e ".[pytorch]" # with PyTorch export support
pip install -e ".[all]"     # everything

You also need a MINLP solver accessible to Pyomo, for example Couenne or SCIP.

Quick start

import optiml
from optiml.losses import MSELoss

# Ultra-small Edge AI classifier — only 9 parameters
model = optiml.Sequential(
    optiml.Linear(2, 2),
    optiml.ReLU(M=10),
    optiml.Linear(2, 1),
)

model.fit(X_train, y_train, loss=MSELoss(reduction='sum'), solver='couenne')

# Export to PyTorch for inference / deployment
pytorch_model = model.export('pytorch')

Available layers

Layer	Description
Linear(in, out)	Fully-connected layer
Conv1D(in_ch, out_ch, kernel)	1-D convolution
Conv2D(in_ch, out_ch, kernel)	2-D convolution
ReLU(M)	ReLU via big-M formulation
AvgPool2D(kernel)	2-D average pooling
Flatten()	Flatten spatial dimensions

Available losses

Loss	Description
MSELoss(reduction)	Mean / sum of squared errors
SSELoss()	Sum of squared errors
MAELoss(reduction)	Mean / sum of absolute errors
HuberLoss(delta)	Smooth L1 loss

Export

After fitting, call model.export('pytorch') to get a torch.nn.Sequential with the optimal weights loaded.

Example

See examples/binary_classification.py for a full working example. It trains a 9-parameter Iris flower classifier (versicolor vs virginica from petal measurements) and compares OptiML with PyTorch + Adam:

• OptiML finds the mathematically optimal weights in ~7 s, achieving 93.3 % test accuracy on 90 unseen samples.
• PyTorch Adam has a 30 % failure rate (stuck at 50 % accuracy) and even the best restart (selected by training loss) only reaches 91.1 %.