MLPClassifier
A dense feed-forward neural network for classification, trained with mini-batch Adam (or SGD) on the cross-entropy loss.
Algorithm
Forward and backward passes are expressed entirely as Eigen matrix operations. The output layer is sigmoid (binary) or softmax (multiclass). Adam moment vectors persist across partial_fit calls for warm-started streaming.
Constructor
Skigen::MLPClassifier<Scalar> model(std::vector<int> hidden_layer_sizes = {100}, Activation = ReLU, Solver = Adam, ...);
Parameters
| Parameter | Default | Description |
|---|---|---|
hidden_layer_sizes | {100} | Units per hidden layer. |
activation | ReLU | Identity, Logistic, Tanh, ReLU. |
solver | Adam | Adam or SGD (LBFGS reserved). |
alpha | 1e-4 | L2 regularisation strength. |
learning_rate_init | 1e-3 | Initial step size. |
max_iter | 200 | Maximum epochs. |
random_state | nullopt | Seed. |
Methods
| Method | Description |
|---|---|
fit(X, y) | Train the network. |
predict(X) | Class labels. |
predict_proba(X) | Softmax/sigmoid outputs. |
score(X, y) | Mean accuracy. |
Fitted Attributes
| Accessor | Description |
|---|---|
coefs() | Per-layer weight matrices. |
intercepts() | Per-layer bias vectors. |
n_iter_run() | Epochs actually run. |
Example
Skigen::MLPClassifier<double> mlp({64, 32});
mlp.fit(X, y);
auto preds = mlp.predict(X_test);
Known Gaps
solver = LBFGS is explicitly feature-gated for v1.2.0. Constructing an MLPClassifier with MLPSolver::LBFGS throws a clear std::invalid_argument; Adam and SGD remain supported.
This estimator is checked by the parity suite. See the generator tests/parity/generate_neural_network_reference.py and the reference fixtures in tests/parity/data/mlp_classifier/, exercised by tests/parity/parity_neural_network.cpp.
For full signatures see the MLPClassifier API Reference.