Skip to content

Latest commit

 

History

History
340 lines (213 loc) · 10.3 KB

concrete.ml.sklearn.qnn.md

File metadata and controls

340 lines (213 loc) · 10.3 KB
Raw

module concrete.ml.sklearn.qnn

Scikit-learn interface for fully-connected quantized neural networks.

Global Variables

  • QNN_AUTO_KWARGS

class NeuralNetRegressor

A Fully-Connected Neural Network regressor with FHE.

This class wraps a quantized neural network implemented using Torch tools as a scikit-learn estimator. The skorch package allows to handle training and scikit-learn compatibility, and adds quantization as well as compilation functionalities. The neural network implemented by this class is a multi layer fully connected network trained with Quantization Aware Training (QAT).

Inputs and targets that are float64 will be casted to float32 before training as Torch does not handle float64 types properly. Thus should not have a significant impact on the model's performances. An error is raised if these values are not floating points.

method __init__

__init__(
    criterion=<class 'torch.nn.modules.loss.MSELoss'>,
    optimizer=<class 'torch.optim.adam.Adam'>,
    lr=0.01,
    max_epochs=10,
    batch_size=128,
    iterator_train=<class 'torch.utils.data.dataloader.DataLoader'>,
    iterator_valid=<class 'torch.utils.data.dataloader.DataLoader'>,
    dataset=<class 'skorch.dataset.Dataset'>,
    train_split=None,
    callbacks=None,
    predict_nonlinearity='auto',
    warm_start=False,
    verbose=1,
    device='cpu',
    **kwargs
)

property base_module

Get the Torch module.

Returns:

  • SparseQuantNeuralNetwork: The fitted underlying module.

property fhe_circuit

property history

property input_quantizers

Get the input quantizers.

Returns:

  • List[UniformQuantizer]: The input quantizers.

property is_compiled

Indicate if the model is compiled.

Returns:

  • bool: If the model is compiled.

property is_fitted

Indicate if the model is fitted.

Returns:

  • bool: If the model is fitted.

property onnx_model

Get the ONNX model.

Is None if the model is not fitted.

Returns:

  • onnx.ModelProto: The ONNX model.

property output_quantizers

Get the output quantizers.

Returns:

  • List[UniformQuantizer]: The output quantizers.

method fit

fit(
    X: Union[ndarray, Tensor, ForwardRef('DataFrame'), List],
    y: Union[ndarray, Tensor, ForwardRef('DataFrame'), ForwardRef('Series'), List],
    *args,
    **kwargs
)

method fit_benchmark

fit_benchmark(
    X: Union[ndarray, Tensor, ForwardRef('DataFrame'), List],
    y: Union[ndarray, Tensor, ForwardRef('DataFrame'), ForwardRef('Series'), List],
    *args,
    **kwargs
)

method predict

predict(
    X: Union[ndarray, Tensor, ForwardRef('DataFrame'), List],
    fhe: Union[FheMode, str] = <FheMode.DISABLE: 'disable'>
) → ndarray

method predict_proba

predict_proba(
    X: Union[ndarray, Tensor, ForwardRef('DataFrame'), List],
    fhe: Union[FheMode, str] = <FheMode.DISABLE: 'disable'>
) → ndarray

class NeuralNetClassifier

A Fully-Connected Neural Network classifier with FHE.

This class wraps a quantized neural network implemented using Torch tools as a scikit-learn estimator. The skorch package allows to handle training and scikit-learn compatibility, and adds quantization as well as compilation functionalities. The neural network implemented by this class is a multi layer fully connected network trained with Quantization Aware Training (QAT).

Inputs that are float64 will be casted to float32 before training as Torch does not handle float64 types properly. Thus should not have a significant impact on the model's performances. If the targets are integers of lower bit-width, they will be safely casted to int64. Else, an error is raised.

method __init__

__init__(
    criterion=<class 'torch.nn.modules.loss.CrossEntropyLoss'>,
    optimizer=<class 'torch.optim.adam.Adam'>,
    classes=None,
    lr=0.01,
    max_epochs=10,
    batch_size=128,
    iterator_train=<class 'torch.utils.data.dataloader.DataLoader'>,
    iterator_valid=<class 'torch.utils.data.dataloader.DataLoader'>,
    dataset=<class 'skorch.dataset.Dataset'>,
    train_split=None,
    callbacks=None,
    predict_nonlinearity='auto',
    warm_start=False,
    verbose=1,
    device='cpu',
    **kwargs
)

property base_module

Get the Torch module.

Returns:

  • SparseQuantNeuralNetwork: The fitted underlying module.

property classes_

property fhe_circuit

property history

property input_quantizers

Get the input quantizers.

Returns:

  • List[UniformQuantizer]: The input quantizers.

property is_compiled

Indicate if the model is compiled.

Returns:

  • bool: If the model is compiled.

property is_fitted

Indicate if the model is fitted.

Returns:

  • bool: If the model is fitted.

property onnx_model

Get the ONNX model.

Is None if the model is not fitted.

Returns:

  • onnx.ModelProto: The ONNX model.

property output_quantizers

Get the output quantizers.

Returns:

  • List[UniformQuantizer]: The output quantizers.

method fit

fit(
    X: Union[ndarray, Tensor, ForwardRef('DataFrame'), List],
    y: Union[ndarray, Tensor, ForwardRef('DataFrame'), ForwardRef('Series'), List],
    *args,
    **kwargs
)

method fit_benchmark

fit_benchmark(
    X: Union[ndarray, Tensor, ForwardRef('DataFrame'), List],
    y: Union[ndarray, Tensor, ForwardRef('DataFrame'), ForwardRef('Series'), List],
    *args,
    **kwargs
)

method predict

predict(
    X: Union[ndarray, Tensor, ForwardRef('DataFrame'), List],
    fhe: Union[FheMode, str] = <FheMode.DISABLE: 'disable'>
) → ndarray

method predict_proba

predict_proba(
    X: Union[ndarray, Tensor, ForwardRef('DataFrame'), List],
    fhe: Union[FheMode, str] = <FheMode.DISABLE: 'disable'>
) → ndarray