---

name: neural-networks description: "Use PyGAD's pure NumPy neural helpers and Keras/Torch adapters for GA-based model optimization." metadata: disco-role: operating disable-model-invocation: true license: BSD 3-Clause

neural-networks

Use this sub-skill when the task asks to evolve neural-network weights with PyGAD: pure NumPy dense networks, pure NumPy CNNs, GANN/GACNN populations, or Keras/PyTorch model adapters.

Route elsewhere

• Generic pygad.GA tuning, custom operators, batch fitness, callbacks, or multi-objective selectors: use genetic-algorithm.
• Built-in benchmark problems or quality indicators: use benchmarks.
• Plotting training/fitness curves or exporting reports after a GA run: use results-and-visuals.

Choose the right neural path

User goal	PyGAD module path	Dependency profile	Notes
Small dense network in NumPy	pygad.nn + pygad.gann	Core PyGAD only	Best for transparent examples and tiny classification/regression tasks.
Small CNN in NumPy	pygad.cnn + pygad.gacnn	Core PyGAD only	CPU-only forward pass; keep fixtures tiny.
Existing Keras model	pygad.kerasga	tensorflow/keras extra	Supports Sequential and Functional models; only trainable layer weights enter the chromosome.
Existing PyTorch model	pygad.torchga	torch extra	Works from state_dict() order; predictions use a copied model so the caller model is preserved.

Operating flow

1. Build or receive the model architecture. Use PyGAD's pure NumPy layer classes, a Keras model, or a torch.nn.Module.
2. Convert model weights into GA chromosomes. Use GANN.population_networks plus population_as_vectors(), GACNN.population_networks, KerasGA.population_weights, or TorchGA.population_weights as initial_population.
3. Write a fitness bridge with the modern PyGAD signature. The fitness function must accept (ga_instance, solution, solution_idx) and return a higher-is-better score. For Keras/Torch, load the candidate solution through the adapter's predict() or restore helper inside fitness; for GANN/GACNN, predict from the matching population model and synchronize matrices in on_generation.
4. Run pygad.GA. Start with few generations, small populations, explicit random_seed, and a simple maximization score before scaling to expensive data.
5. Copy the best solution back into the model. Convert the best vector with the matching matrix/dict helper and update the model or use the module's predict() helper.
6. Inspect and export results elsewhere. For plots/reports, route to results-and-visuals after the GA is complete.

Reference map

• API signatures, model-weight conversion rules, optional dependency boundaries, and data contracts: references/api-reference.md
• End-to-end dense/CNN/Keras/Torch recipes and decision points: references/workflows.md
• Common dependency, shape, label, and weight-conversion failures: references/troubleshooting.md

Bundled safe scripts

• scripts/neural_internal_smoke.py: verifies pure NumPy nn, gann, cnn, and gacnn conversion paths with tiny deterministic data.
• scripts/keras_torch_templates.py: lazy-imports either Keras/TensorFlow or PyTorch and runs a tiny CPU GA template; it exits clearly when the optional framework is missing.

Safety notes

• GA-based neural training is expensive compared with gradient descent. Keep initial checks tiny and scale only when the user accepts runtime cost.
• Classification labels used by PyGAD's pure NumPy examples should be integers from 0 to num_classes - 1.
• Keras/Torch adapters are optional; do not install deep-learning frameworks unless the user's task requires them.
• GPU is not required for these helpers. If a user asks for CUDA/MPS/ROCm, verify the framework backend separately before promising accelerator execution.