Module equilibrium-propagation.lib.cost

Classes

class CEnergy (beta)

Abstract base class for all supervised learning cost functions.

Attributes

beta

Scalar weighting factor of the cost function

target

Current target tensor for the cost function

Expand source code

class CEnergy(abc.ABC):
    """
    Abstract base class for all supervised learning cost functions.

    Attributes:
        beta: Scalar weighting factor of the cost function
        target: Current target tensor for the cost function
    """
    def __init__(self, beta):
        super(CEnergy, self).__init__()
        self.beta = beta
        self.target = None

    @abc.abstractmethod
    def compute_energy(self, u_last):
        """
        Compute energy/loss given a prediction.

        Args:
            Activation of the last layer, i.e. the prediction
        """
        return

    def set_target(self, target):
        """
        Set new target tensor for the cost function.

        Args:
            target: target tensor
        """
        self.target = target

Ancestors

• abc.ABC

Subclasses

• CrossEntropy
• SquaredError

Methods

def compute_energy(self, u_last)

Compute energy/loss given a prediction.

Args

Activation of the last layer, i.e. the prediction

Expand source code

@abc.abstractmethod
def compute_energy(self, u_last):
    """
    Compute energy/loss given a prediction.

    Args:
        Activation of the last layer, i.e. the prediction
    """
    return

def set_target(self, target)

Set new target tensor for the cost function.

Args

target

target tensor

Expand source code

def set_target(self, target):
    """
    Set new target tensor for the cost function.

    Args:
        target: target tensor
    """
    self.target = target

class CrossEntropy (beta)

Cross entropy cost function.

Expand source code

class CrossEntropy(CEnergy):
    """
    Cross entropy cost function.
    """
    def __init__(self, beta):
        super(CrossEntropy, self).__init__(beta)

    def compute_energy(self, u_last):
        """
        Compute cross-entropy loss given a prediction.

        Args:
            Activation of the last layer, i.e. the prediction
        """
        loss = torch.nn.functional.cross_entropy(u_last, self.target, reduction='none')
        return self.beta * loss

    def set_target(self, target):
        if target is None:
            self.target = None
        else:
            # Need to transform target for the F.cross_entropy function
            self.target = target.argmax(dim=1)

Ancestors

• CEnergy
• abc.ABC

Methods

def compute_energy(self, u_last)

Compute cross-entropy loss given a prediction.

Args

Activation of the last layer, i.e. the prediction

Expand source code

def compute_energy(self, u_last):
    """
    Compute cross-entropy loss given a prediction.

    Args:
        Activation of the last layer, i.e. the prediction
    """
    loss = torch.nn.functional.cross_entropy(u_last, self.target, reduction='none')
    return self.beta * loss

Inherited members

• CEnergy:
  • set_target

class SquaredError (beta)

Squared energy cost function.

Expand source code

class SquaredError(CEnergy):
    """
    Squared energy cost function.
    """
    def __init__(self, beta):
        super(SquaredError, self).__init__(beta)

    def compute_energy(self, u_last):
        """
        Compute mean squared error loss given a prediction.

        Args:
            Activation of the last layer, i.e. the prediction
        """
        loss = torch.nn.functional.mse_loss(u_last, self.target, reduction='none')
        return self.beta * 1/2 * torch.sum(loss, dim=1)

Ancestors

• CEnergy
• abc.ABC

Methods

def compute_energy(self, u_last)

Compute mean squared error loss given a prediction.

Args

Activation of the last layer, i.e. the prediction

Expand source code

def compute_energy(self, u_last):
    """
    Compute mean squared error loss given a prediction.

    Args:
        Activation of the last layer, i.e. the prediction
    """
    loss = torch.nn.functional.mse_loss(u_last, self.target, reduction='none')
    return self.beta * 1/2 * torch.sum(loss, dim=1)

Inherited members

• CEnergy:
  • set_target