OpenNN - Open Neural Networks Library Files

OpenNN now numbers its major versions by year: 26.0 follows 8.x, and there is no 9.x release.

A new licence

OpenNN 26.0 and later are available under the GNU Affero General Public License, version 3 only, with additional permissions for linking with NVIDIA CUDA, cuDNN and Intel oneMKL and for the code and model files that OpenNN exports. Alternatively, they are available under a commercial licence from Artelnics (artelnics@artelnics.com). OpenNN 8.x and earlier keep the LGPL 2.1 or later. See LICENSING.md (github.com). Contributions are accepted under the Contributor License Agreement.

Highlights

• Module layout and shorter names. The code now lives under opennn/<module>/, with shorter public names: Network, Training, Evaluation, Adam, SGD, LSTM, LevenbergMarquardt, QuasiNewton, Autoencoder, Yolo and InputSelection.
• Four dataset classes replace seven.
• TabularDataset includes forecasting windows.
• TextDataset covers classification, translation, next-token prediction and token/attention-mask inputs.
• Levenberg-Marquardt evaluates damped steps against an immutable curvature matrix and commits only accepted candidate metrics.
• Stricter model loading. These cases are now errors:
• missing parameter files;
• parameter-count mismatches;
• non-finite embedded weights.

The JSON parser is strict, bounded and fuzz-tested.

• Response optimization
• Accepts original variable names in backticks.
• Adds feasibility controls: set_feasibility_evaluations(), set_feasibility_rounds(), set_sampling_budget_multiplier() and set_maximum_consecutive_failures().
• JavaScript export
• Sanitizes feature identifiers.
• Escapes HTML labels.
• Uses a dropdown for categorical inputs.
• Configurable library logging.
• Build and packaging
• NATIVE and PORTABLE CPU targets.
• Shared-library builds.
• A relocatable CMake package.
• A core package without the vision components.
• Reliability. Many CPU and CUDA fixes in buffer ordering, handle release, mixed-precision layouts and dataset device copies.

Migrating from 8.x

26.0 is a major source and model-format migration:

• Recompile consumers. Replacing an 8.x shared library in place is not supported.
• Request find_package(OpenNN 26.0 ...).
• There are no old-name aliases, and former JSON root and factory names are rejected.

The guide to headers, renamed classes and saved names is in AGENTS.md, "26.0 release notes and [migration"](https://github.com/Artelnics/opennn/blob/v26.0.0/AGENTS.md#260.