losses module¶

class losses.CategoricalCrossEntropyLoss(loss_smoother)¶

Bases: losses.Loss

Categorical cross-entropy loss. Usually preceeded by a linear activation. For multi-class classification. Inherits everything from class Loss.

cache¶

Run-time cache of attibutes such as gradients.

Type: dict

loss_smoother¶

Loss smoother function.

Type: LossSmoother

name¶

The name of the loss.

Type: str

repr_str¶

The string representation of the loss.

Type: str

__init__()¶: Constuctor.

compute_loss(scores, y, layers_reg_loss)¶: Computes loss of classifier - also includes the regularization losses from previous layers.

grad()¶: Computes the gradient of the loss function.

compute_loss(scores, y)¶

Computes loss of classifier - also includes the regularization losses from previous layers.

Parameters

scores (numpy.ndarray) – Scores. Usually from softmax activation. Shape is (batch size, number of classes)
y (numpy.ndarray) – True labels. Shape is (batch size, )

Returns

loss – The overall loss of the classifier.

Return type

float

Notes

None

grad()¶

Computes the gradient of the loss function.

Parameters: None –
Returns: None if gradient has not yet been computed. Shape of gradient is (batch size, ). Note that the grad here is just y.
Return type: numpy.ndarray or None

Notes

None

class losses.CategoricalHingeLoss(loss_smoother)¶

Bases: losses.Loss

Categorical Hinge loss for realizing an SVM classifier. Usually preceeded by a linear activation. For multi-class classification. Inherits everything from class Loss.

cache¶

Run-time cache of attibutes such as gradients.

Type: dict

loss_smoother¶

Loss smoother function.

Type: LossSmoother

name¶

The name of the loss.

Type: str

repr_str¶

The string representation of the loss.

Type: str

__init__()¶: Constuctor.

compute_loss(scores, y, layers_reg_loss)¶: Computes loss of classifier - also includes the regularization losses from previous layers.

grad()¶: Computes the gradient of the loss function.

compute_loss(scores, y)¶

Computes loss of classifier - also includes the regularization losses from previous layers.

Parameters

scores (numpy.ndarray) – Scores. Usually from linear activation. Shape is (number of data points, number of classes)
y (numpy.ndarray) – True labels. Shape is (number of data points, )

Returns

loss – The overall loss of the classifier.

Return type

float

Notes

None

grad()¶

Computes the gradient of the loss function.

Parameters: None –
Returns: None if gradient has not yet been computed. Shape of gradient is (batch size, number of classes)
Return type: numpy.ndarray or None

Notes

None

class losses.Loss(name, loss_smoother)¶

Bases: object

Loss parent class.

cache¶

Run-time cache of attributes such as gradients.

Type: dict

__init__()¶: Constuctor.

class losses.LossSmoother(repr_str)¶

Bases: object

Loss smoother parent class. Smooths out abruptly varying losses.

first_call¶

If the loss smoother is called for the first time.

Type: bool

repr_str¶

The string representation of the loss smoother.

Type: str

__init__()¶: Constuctor.

__call__(loss)¶: Computes the smoothed loss.

class losses.LossSmootherConstant¶

Bases: losses.LossSmoother

Constant loss smoother. Does not smooth the loss.

first_call¶

If the loss smoother is called for the first time.

Type: bool

repr_str¶

The string representation of the loss smoother.

Type: str

cache¶

Run-time cache of variables needed for computing the smoothed loss.

Type: dict

__init__()¶: Constuctor.

__call__(loss)¶: Computes the smoothed loss.

class losses.LossSmootherMovingAverage(alpha)¶

Bases: losses.LossSmoother

Moving average loss smoother.

first_call¶

If the loss smoother is called for the first time.

Type: bool

repr_str¶

The string representation of the loss smoother.

Type: str

cache¶

Run-time cache of variables needed for computing the smoothed loss.

Type: dict

__init__()¶: Constuctor.

__call__(loss)¶: Computes the smoothed loss.

class losses.MeanSquaredErrorLoss(loss_smoother)¶

Bases: losses.Loss

MSE loss.

cache¶

Run-time cache of attibutes such as gradients.

Type: dict

loss_smoother¶

Loss smoother function.

Type: LossSmoother

name¶

The name of the loss.

Type: str

repr_str¶

The string representation of the loss.

Type: str

__init__()¶: Constuctor.

compute_loss(scores, y, layers_reg_loss)¶: Computes loss of classifier.

grad()¶: Computes the gradient of the loss function.

compute_loss(scores, y)¶

Computes loss of classifier - also includes the regularization losses from previous layers.

Parameters

scores (numpy.ndarray) – Scores. Usually from softmax activation. Shape is (batch size, )
y (numpy.ndarray) – True labels. Shape is (batch size, )

Returns

loss – The overall loss of the classifier.

Return type

float

Notes

None

grad()¶

Computes the gradient of the loss function.

Parameters: None –
Returns: None if gradient has not yet been computed. Shape of gradient is (batch size, ). Note that the grad here is just y.
Return type: numpy.ndarray or None

Notes

None