module `training.sgd_learning_loss`#

Short summary#

module onnxcustom.training.sgd_learning_loss

Classes#

class	truncated documentation
`AbsoluteLearningLoss`	Implements a square loss $\|Y - Z\|$ where Y is the output and Z the expected output. See `_onnx_grad_loss_absolute_error()` …
`BaseLearningLoss`	Class handling the loss for class `OrtGradientForwardBackwardOptimizer`. All classes inheriting from this …
`ElasticLearningLoss`	Implements a square loss $(Y - Z)^2 \alpha + \|Y - Z\| * \beta$ where Y is the output and Z the expected …
`NegLogLearningLoss`	Implements a negative log loss ‘log(yt, yp) = -(1-yt)log(1-yp) - ytlog(yp), this only works for a binary classification …
`SquareLearningLoss`	Implements a square loss $(Y - Z)^2$ where Y is the output and Z the expected output. See `_onnx_grad_loss_square_error()` …

Static Methods#

staticmethod	truncated documentation
`select`	Returns an instance of a given initialized with kwargs.
`select`	Returns an instance of a given initialized with kwargs.
`select`	Returns an instance of a given initialized with kwargs.
`select`	Returns an instance of a given initialized with kwargs.
`select`	Returns an instance of a given initialized with kwargs.

Methods#

method	truncated documentation
`__init__`
`__init__`
`__init__`
`__init__`
`__init__`
`_call_iobinding`
`_call_iobinding`
`_call_iobinding`
`_call_iobinding`
`_call_iobinding`
`build_onnx_function`
`build_onnx_function`
`build_onnx_function`
`build_onnx_function`
`build_onnx_score_function`	Assuming the loss function was created. This one takes the onnx graph and generate the onnx graph for the …
`build_onnx_score_function`	Assuming the loss function was created. This one takes the onnx graph and generate the onnx graph for the …
`build_onnx_score_function`	Assuming the loss function was created. This one takes the onnx graph and generate the onnx graph for the …
`build_onnx_score_function`	Assuming the loss function was created. This one takes the onnx graph and generate the onnx graph for the …
`build_onnx_score_function`	Assuming the loss function was created. This one takes the onnx graph and generate the onnx graph for the …
`loss_gradient`	Returns the loss and the gradient as OrtValue.
`loss_gradient`	Returns the loss and the gradient as OrtValue.
`loss_gradient`	Returns the loss and the gradient as OrtValue.
`loss_gradient`	Returns the loss and the gradient as OrtValue.
`loss_gradient`	Returns the loss and the gradient as OrtValue.
`loss_scores`	Returns the weighted loss (or score) for every observation as OrtValue.
`loss_scores`	Returns the weighted loss (or score) for every observation as OrtValue.
`loss_scores`	Returns the weighted loss (or score) for every observation as OrtValue.
`loss_scores`	Returns the weighted loss (or score) for every observation as OrtValue.
`loss_scores`	Returns the weighted loss (or score) for every observation as OrtValue.

Documentation#

Helper for onnxruntime-training.

source on GitHub

class onnxcustom.training.sgd_learning_loss.AbsoluteLearningLoss#

Bases: BaseLearningLoss

Implements a square loss $|Y - Z|$ where Y is the output and Z the expected output. See _onnx_grad_loss_absolute_error for the ONNX implementation.

source on GitHub

__init__()#

build_onnx_function(opset, device, weight_name)#

This class computes a function represented as an ONNX graph. This method builds it. This function creates InferenceSession which do that.

Parameters:

opset – opset to use
device – C_OrtDevice
args – additional arguments

source on GitHub

class onnxcustom.training.sgd_learning_loss.BaseLearningLoss#

Bases: BaseLearningOnnx

Class handling the loss for class OrtGradientForwardBackwardOptimizer. All classes inheriting from this one creates one ONNX function, returning the loss and the gradient of the loss against the outputs. Method loss_gradient is the main method, it computes the loss and the gradient defiend by one ONNX graph and executed by an instance of InferenceSession.

source on GitHub

__init__()#

_call_iobinding(sess, bind)#

build_onnx_score_function(opset, device, weight_name)#

Assuming the loss function was created. This one takes the onnx graph and generate the onnx graph for the method loss_score.

source on GitHub

loss_gradient(device, expected, predicted, weight=None)#

Returns the loss and the gradient as OrtValue.

Parameters:

device – device where the training takes place
expected – expected value
predicted – predicted value
weight – optional, training weights (same dimension as expected and predicted tensors)

Returns:

loss and gradient

source on GitHub

loss_scores(device, expected, predicted, weight=None)#

Returns the weighted loss (or score) for every observation as OrtValue.

Parameters:

device – device where the training takes place
expected – expected value
predicted – predicted value
weight – optional, training weights (same dimension as expected and predicted tensors)

Returns:

a score for every observation

source on GitHub

static select(class_name, **kwargs)#

Returns an instance of a given initialized with kwargs. :param class_name: an instance of BaseLearningLoss

or a string among the following class names (see below)

Returns:: instance of BaseLearningLoss

Possible values for class_name: * ‘square_error’: see SquareLearningLoss * ‘absolute_error’: see AbsoluteLearningLoss * ‘elastic_error’: see ElasticLearningLoss

source on GitHub

class onnxcustom.training.sgd_learning_loss.ElasticLearningLoss(l1_weight=0.5, l2_weight=0.5)#

Bases: BaseLearningLoss

Implements a square loss $(Y - Z)^2 \alpha + |Y - Z| * \beta$ where Y is the output and Z the expected output, $\alpha$ is l2_weight and $\beta$ is l1_weight.

Parameters:

l1_weight – weight of L1 norm
l2_weight – weight of L2 norm

See _onnx_grad_loss_elastic_error for the ONNX implementation.

source on GitHub

__init__(l1_weight=0.5, l2_weight=0.5)#

build_onnx_function(opset, device, weight_name)#

This class computes a function represented as an ONNX graph. This method builds it. This function creates InferenceSession which do that.

Parameters:

opset – opset to use
device – C_OrtDevice
args – additional arguments

source on GitHub

class onnxcustom.training.sgd_learning_loss.NegLogLearningLoss(eps=1e-05, probability_function='sigmoid')#

Bases: BaseLearningLoss

Implements a negative log loss ‘log(yt, yp) = -(1-yt)log(1-yp) - ytlog(yp), this only works for a binary classification where yp is the predicted probability, yt is the expected probability. yt is expected to be binary, yp is a matrix with two columns, the sum on every line is 1. However, this loss is usually applied after a function softmax and the gradient is directly computed from the loss to the raw score before they are processed through the softmax function (see class Log).

Parameters:

eps – clipping value for probabilities, avoids computing log(0)
probability_function – function to convert raw scores into probabilities, default value is sigmoid for a logistic regression

source on GitHub