pymoose.predictors package

Submodules

• pymoose.predictors.linear_predictor module
• pymoose.predictors.multilayer_perceptron_predictor module
• pymoose.predictors.neural_network_predictor module
• pymoose.predictors.onnx_convert module
• pymoose.predictors.predictor module
• pymoose.predictors.predictor_utils module
• pymoose.predictors.tree_ensemble module

Module contents

class pymoose.predictors.LinearClassifier(coeffs, intercepts=None, post_transform=None)

Bases: pymoose.predictors.linear_predictor.LinearPredictor

Linear classifier predictor interface.

Parameters

• coeffs – Array-like convertible to a (n_outputs, n_weights)-shaped ndarray.

• intercepts – Optional array-like convertible to a vector.

• post_transform – a PostTransform enum variant describing how to convert the raw linear model scores into probabilistic classification outputs.

classmethod from_onnx(model_proto)

Construct LinearClassifier from a parsed ONNX model.

Parameters

model_proto – An ONNX ModelProto containing a LinearClassifier operator node.

Returns

A LinearClassifier with parameters and model configuration loaded from the ONNX model.

Raises

• ValueError if ONNX graph is missing expected nodes. –

• RuntimeError if ONNX LinearClassifier node has an unsupported post-transform – function attribute.

post_transform(y)

Applies the classifier’s post transform function to a tensor.

Actual function used depends on value of PostTransform and number of classes provided at instantiation.

Parameters

y – A tensor

Returns

A tensor.

class pymoose.predictors.LinearRegressor(coeffs, intercepts=None)

Bases: pymoose.predictors.linear_predictor.LinearPredictor

Linear regression predictor interface.

Parameters

• coeffs – Array-like convertible to a (n_outputs, n_weights)-shaped ndarray.

• intercepts – Optional array-like convertible to a vector.

classmethod from_onnx(model_proto)

Construct LinearRegressor from a parsed ONNX model.

Parameters

model_proto – An ONNX ModelProto containing a LinearRegressor operator node.

Returns

A LinearRegressor with weighhts and bias terms loaded from the ONNX model.

Raises

ValueError if ONNX graph is missing expected nodes. –

post_transform(y)

Applies no-op to linear predictor function output.

class pymoose.predictors.MLPClassifier(weights, biases, activation)

Bases: pymoose.predictors.multilayer_perceptron_predictor.MLPPredictor

post_transform(y, fixedpoint_dtype)

class pymoose.predictors.MLPRegressor(weights, biases, activation)

Bases: pymoose.predictors.multilayer_perceptron_predictor.MLPPredictor

post_transform(y, fixedpoint_dtype)

class pymoose.predictors.NeuralNetwork(weights, biases, activations)

Bases: pymoose.predictors.predictor.Predictor

activation_fn(z, i)

apply_layer(input, i, fixedpoint_dtype)

classmethod from_onnx(model_proto)

predictor_fn(x, fixedpoint_dtype)

class pymoose.predictors.TreeEnsembleClassifier(trees, n_features, n_classes, base_score, learning_rate, transform_output, tree_class_map)

Bases: pymoose.predictors.tree_ensemble.TreeEnsemble

Tree ensemble classification Predictor for GBTs and Random Forests.

This class can be used for binary, multiclass, or multilabel classification. Support for multiclass classification uses the one-vs-rest method.

Parameters

• trees – Nested collection of :class:`~DecisionTreeRegressor`s.

• n_features – Number of features expected for input data.

• n_classes – Number of output classes.

• base_score – The base score for the underlying tree ensemble model, similar to a bias/intercept term.

• learning_rate – Learning rate parameter used to re-scale leaf weights in the model trees.

• transform_output – Boolean determining whether a softmax should be applied to derive probabilities from the tree ensemble output.

• tree_class_map – Dictionary mapping trees indices to class indices. Keeps track of which trees in trees correspond to each class in the one-vs-rest formulation of multiclass classification.

classmethod from_onnx(model_proto)

Construct a TreeEnsembleClassifier from a parsed ONNX model.

Parameters

model_proto – An ONNX ModelProto containing a TreeEnsembleClassifier node.

Returns

A TreeEnsembleClassifier built from the parameters and configuration of the given ONNX model.

Raises

ValueError if ONNX graph is missing expected nodes. –

post_transform(tree_scores, fixedpoint_dtype)

class pymoose.predictors.TreeEnsembleRegressor(trees, n_features, base_score, learning_rate)

Bases: pymoose.predictors.tree_ensemble.TreeEnsemble

Tree ensemble regression Predictor, accommodating both GBTs and Random Forests.

Parameters

• trees – Nested collection of :class:`~DecisionTreeRegressor`s.

• n_features – Number of features expected for input data.

• base_score – The base score for the underlying tree ensemble model, similar to a bias/intercept term.

• learning_rate – Learning rate parameter used to re-scale leaf weights in the model trees.

classmethod from_onnx(model_proto)

Construct a TreeEnsembleRegressor from a parsed ONNX model.

Parameters

model_proto – An ONNX ModelProto containing a TreeEnsembleRegressor node.

Returns

A TreeEnsembleRegressor built from the parameters and configuration of the given ONNX model.

Raises

ValueError if ONNX graph is missing expected nodes. –

post_transform(tree_scores, fixedpoint_dtype)

pymoose.predictors.from_onnx(model_proto)

Attempt to infer and construct a Moose predictor type from a given ONNX Model.

Parameters

model_proto – An ONNX ModelProto containing nodes sufficient to construct some Moose predictor.

Returns

A Moose predictor, inferred from the structure and contents of the ONNX model.

Raises

• ValueError if the Predictor type cannot be inferred, or if the ONNX graph is – malformed for the inferred Predictor type.

• RuntimeError if LinearClassifier constructor throws an error due to unrecognized – post_transform attribute.