Module equilibrium-propagation.lib.utils

Functions

def create_activations(name, n_layers)

Create activation functions for every layer of the network.

Args

name

Name of the activation function

n_layers

Number of layers

Returns

List of activation functions for every layer

 

Expand source code

def create_activations(name, n_layers):
    """
    Create  activation functions for every layer of the network.

    Args:
        name: Name of the activation function
        n_layers: Number of layers

    Returns:
        List of activation functions for every layer
    """
    if name == 'relu':
        phi_l = torch.relu
    elif name == "leaky_relu":
        def phi_l(x): torch.nn.functional.leaky_relu(x, negative_slope=0.05)
    elif name == 'softplus':
        phi_l = torch.nn.functional.softplus
    elif name == 'sigmoid':
        phi_l = torch.sigmoid
    elif name == 'hard_sigmoid':
        def phi_l(x): torch.clamp(x, min=0, max=1)
    else:
        raise ValueError(f'Nonlinearity \"{name}\" not defined.')

    return [lambda x: x] + [phi_l] * (n_layers - 1)

def create_cost(name, beta)

Create a supervised learning cost function used to nudge the network towards a desired state during training.

Args

name

Name of the cost function

beta

Scalar weighting factor of the cost function

Returns

CEnergy object

 

Expand source code

def create_cost(name, beta):
    """
    Create a supervised learning cost function used to nudge
    the network towards a desired state during training.

    Args:
        name: Name of the cost function
        beta: Scalar weighting factor of the cost function

    Returns:
        CEnergy object
    """
    if name == "squared_error":
        return cost.SquaredError(beta)
    elif name == "cross_entropy":
        return cost.CrossEntropy(beta)
    else:
        raise ValueError("Cost function \"{}\" not defined".format(name))

def create_optimizer(model, name, **kwargs)

Create optimizer for the given model.

Args

model

nn.Module whose parameters will be optimized

name

Name of the optimizer to be used

Returns

torch.optim.Optimizer instance for the given model

 

Expand source code

def create_optimizer(model, name, **kwargs):
    """
    Create optimizer for the given model.

    Args:
        model: nn.Module whose parameters will be optimized
        name: Name of the optimizer to be used

    Returns:
        torch.optim.Optimizer instance for the given model
    """
    if name == "adagrad":
        return torch.optim.Adagrad(model.parameters(), **kwargs)
    elif name == "adam":
        return torch.optim.Adam(model.parameters(), **kwargs)
    elif name == "sgd":
        return torch.optim.SGD(model.parameters(), **kwargs)
    else:
        raise ValueError("Optimizer \"{}\" undefined".format(name))