Metrics

tn4ml.metrics.NegLogLikelihood(model, data)

Negative Log-Likelihood loss.

Parameters:

• model (quimb.tensor.MatrixProductState) – Matrix Product State model

• data (quimb.tensor.MatrixProductState) – Input MPS

Return type:

float

tn4ml.metrics.TransformedSquaredNorm(model, data)

Squared norm of transformed input data.

Parameters:

• model (tn4ml.models.smpo.SpacedMatrixProductOperator) – Spaced Matrix Product Operator

• data (quimb.tensor.MatrixProductState) – Input mps.

Return type:

float

tn4ml.metrics.NoReg(x)

tn4ml.metrics.LogFrobNorm(model)

Regularization cost - log(Frobenius-norm of model)

Parameters:

model (quimb.tensor.MatrixProductState) – Matrix Product State model

Return type:

float

tn4ml.metrics.LogPowFrobNorm(model)

Regularization cost - log(squared(Frobenius-norm of model))

Parameters:

model (quimb.tensor.MatrixProductState) – Matrix Product State model

Return type:

float

tn4ml.metrics.LogReLUFrobNorm(model)

Regularization cost using ReLU of the log of the Frobenius-norm of model.

Parameters:

model (tn4ml.models.smpo.SpacedMatrixProductOperator) – Spaced Matrix Product Operator

Return type:

float

tn4ml.metrics.QuadFrobNorm(model)

Regularization cost using the quadratic formula centered in 1 of the Frobenius-norm of model.

Parameters:

model (tn4ml.models.smpo.SpacedMatrixProductOperator) – Spaced Matrix Product Operator

Return type:

float

tn4ml.metrics.LogQuadNorm(model, data)

Example of error calculation when applying tn4ml.models.smpo.SpacedMatrixProductOperator P to data.

Parameters:

• model (tn4ml.models.smpo.SpacedMatrixProductOperator) – Spaced Matrix Product Operator

• data (quimb.tensor.MatrixProductState) – Input mps.

Return type:

float

tn4ml.metrics.QuadNorm(model, data)

Example of error calculation when applying tn4ml.models.smpo.SpacedMatrixProductOperator P to data.

Parameters:

• model (tn4ml.models.smpo.SpacedMatrixProductOperator) – Spaced Matrix Product Operator

• data (quimb.tensor.MatrixProductState) – Input mps.

Return type:

float

tn4ml.metrics.SemiSupervisedLoss(model, data, y_true, **kwargs)

Loss function for semi-supervised learning.

Parameters:

• model (tn4ml.models.smpo.SpacedMatrixProductOperator) – Spaced Matrix Product Operator

• data (quimb.tensor.MatrixProductState) – Input Matrix Product State

• y_true (Number) – Target class percentage.

Return type:

float

tn4ml.metrics.SemiSupervisedNLL(model, data, y_true=None, **kwargs)

Loss function for semi-supervised learning.

Parameters:

• model (tn4ml.models.smpo.SpacedMatrixProductOperator) – Spaced Matrix Product Operator

• data (quimb.tensor.MatrixProductState) – Input Matrix Product State

• y_true (Number) – Target class percentage.

Return type:

float

tn4ml.metrics.Softmax(z, position)

Softmax function.

Parameters:

• z (jnp.array`) – Predicted probabilities.

• position (int) – Indicates for which class we are calculating softmax value.

Return type:

float

tn4ml.metrics.CrossEntropySoftmax(model, data, targets)

Cross-entropy loss function for supervised learning.

Parameters:

• model (tn4ml.models.smpo.SpacedMatrixProductOperator) – Spaced Matrix Product Operator

• data (quimb.tensor.MatrixProductState) – Input Matrix Product State

• targets (numpy.ndarray) – Target class vector. Example = [1 0 0 0] for n_classes = 4.

Return type:

float

tn4ml.metrics.MeanSquaredError(model, data, targets)

Mean Squared Error loss function for supervised learning.

Parameters:

• model (tn4ml.models.smpo.SpacedMatrixProductOperator) – Spaced Matrix Product Operator

• data (quimb.tensor.MatrixProductState) – Input Matrix Product State

• targets (numpy.ndarray) – Target class vector. Example = [1 0 0 0] for n_classes = 4.

Return type:

float

tn4ml.metrics.OptaxWrapper(optax_loss=None)

Wrapper around optax loss functions for supervised learning. Make sure you got all inputs to loss function correct. Refer to documentation for each loss to https://optax.readthedocs.io/en/latest/api/losses.html . Make sure SMPO has only one output with dimension = number of classes.

Parameters:

• model (tn4ml.models.model.Model) – Tensor Network model.

• data (quimb.tensor.MatrixProductState) – Input Matrix Product State

• y_true (numpy.ndarray) – Target class vector. Example = [1 0 0 0] for n_classes = 4.

• kwargs (dict) – Additional arguments for optax loss function.

Return type:

float

tn4ml.metrics.CrossEntropyWeighted(class_weights=None)

Parameters:

class_weights (array)

Return type:

Callable

tn4ml.metrics.CombinedLoss(model, data, y_true=None, error=<function LogQuadNorm>, reg=<function NoReg>, embedding=None)

Unified Loss function combining error computation and regularization.

Parameters:

• model (tn4ml.models.Model) – Tensor Network with parametrized tensors.

• data (Union[qtn.MatrixProductState, numpy.ndarray]) – Data used for computing the loss value.

• y_true (Optional[jnp.array]) – True labels for supervised learning tasks (only applicable for MatrixProductState data).

• error (function) – Function for error calculation.

• reg (function) – Function for regularization value calculation.

• embedding (tn4ml.embeddings.Embedding, optional) – Data embedding function.

Returns:

Computed loss value.

Return type:

float