yobx.sklearn.neural_network.mlp#

yobx.sklearn.neural_network.mlp.sklearn_mlp_classifier(g: GraphBuilderExtendedProtocol, sts: Dict, outputs: List[str], estimator: MLPClassifier, X: str, name: str = 'mlp_classifier') → Tuple[str, str][source]#

Converts a sklearn.neural_network.MLPClassifier into ONNX.

All hidden layers use the activation stored in estimator.activation; the output layer uses estimator.out_activation_ which is 'logistic' for binary classification and 'softmax' for multi-class.

Hidden layers (repeated for each weight matrix except the last):

h_prev  ──MatMul(coef_i)──Add(bias_i)──Activation──►  h_i

Binary classification (out_activation_ == 'logistic'):

h  ──MatMul(coef_out)──Add(bias_out)──Sigmoid──►  proba_pos
                                                       │
                                      Sub(1, ·) ──►  proba_neg
                                             │
                                      Concat  ──►  probabilities
                                        │
                                   ArgMax──Cast──Gather(classes)──►  label

Multi-class classification (out_activation_ == 'softmax'):

h  ──MatMul(coef_out)──Add(bias_out)──Softmax──►  probabilities
                                                       │
                                   ArgMax──Cast──Gather(classes)──►  label

Parameters:

g – the graph builder to add nodes to
sts – shapes defined by scikit-learn
estimator – a fitted MLPClassifier
outputs – desired names (label, probabilities)
X – input tensor name
name – prefix names for the added nodes

Returns:

tuple (label_result_name, proba_result_name)

yobx.sklearn.neural_network.mlp.sklearn_mlp_regressor(g: GraphBuilderExtendedProtocol, sts: Dict, outputs: List[str], estimator: MLPRegressor, X: str, name: str = 'mlp_regressor') → str[source]#

Converts a sklearn.neural_network.MLPRegressor into ONNX.

All hidden layers use the activation stored in estimator.activation ('identity', 'logistic', 'tanh', or 'relu'); the output layer always uses the 'identity' activation (i.e., the linear output is returned as-is).

X  ──hidden layers──►  h  ──MatMul(coef_out)──Add(bias_out)──►  predictions

Parameters:

g – the graph builder to add nodes to
sts – shapes defined by scikit-learn
estimator – a fitted MLPRegressor
outputs – desired output names (predictions)
X – input tensor name
name – prefix names for the added nodes

Returns:

output tensor name

yobx.sklearn.neural_network.mlp#

This Page