ceml.backend.torch

ceml.backend.torch.costfunctions

class ceml.backend.torch.costfunctions.costfunctions.CostFunctionDifferentiableTorch(**kwds)

Bases: ceml.costfunctions.costfunctions.CostFunctionDifferentiable

Base class of differentiable cost functions implemented in PyTorch.

grad()

Warning

Do not use this method!

Call ‘.backward()’ of the output tensor. After that, the gradient of each variable ‘myvar’ - that is supposed to have gradient - can be accessed as ‘myvar.grad’

Raises

NotImplementedError –

class ceml.backend.torch.costfunctions.costfunctions.DummyCost(**kwds)

Bases: ceml.backend.torch.costfunctions.costfunctions.CostFunctionDifferentiableTorch

Dummy cost function - always returns zero.

score_impl(x)

Computes the loss - always return zero.

class ceml.backend.torch.costfunctions.costfunctions.L1Cost(x_orig, **kwds)

Bases: ceml.backend.torch.costfunctions.costfunctions.CostFunctionDifferentiableTorch

L1 cost function.

score_impl(x)

Computes the loss - l1 norm.

class ceml.backend.torch.costfunctions.costfunctions.L2Cost(x_orig, **kwds)

Bases: ceml.backend.torch.costfunctions.costfunctions.CostFunctionDifferentiableTorch

L2 cost function.

score_impl(x)

Computes the loss - l2 norm.

class ceml.backend.torch.costfunctions.costfunctions.LMadCost(x_orig, mad, **kwds)

Bases: ceml.backend.torch.costfunctions.costfunctions.CostFunctionDifferentiableTorch

Manhattan distance weighted feature-wise with the inverse median absolute deviation (MAD).

score_impl(x)

Computes the loss.

class ceml.backend.torch.costfunctions.costfunctions.MinOfListCost(dist, samples, **kwds)

Bases: ceml.backend.torch.costfunctions.costfunctions.CostFunctionDifferentiableTorch

Minimum distance to a list of data points.

score_impl(x)

Computes the loss.

class ceml.backend.torch.costfunctions.costfunctions.NegLogLikelihoodCost(y_target, **kwds)

Bases: ceml.backend.torch.costfunctions.costfunctions.CostFunctionDifferentiableTorch

Negative-log-likelihood cost function.

score_impl(y)

Computes the loss - negative-log-likelihood.

class ceml.backend.torch.costfunctions.costfunctions.RegularizedCost(penalize_input, penalize_output, C=1.0, **kwds)

Bases: ceml.backend.torch.costfunctions.costfunctions.CostFunctionDifferentiableTorch

Regularized cost function.

score_impl(x)

Computes the loss.

class ceml.backend.torch.costfunctions.costfunctions.SquaredError(y_target, **kwds)

Bases: ceml.backend.torch.costfunctions.costfunctions.CostFunctionDifferentiableTorch

Squared error cost function.

score_impl(y)

Computes the loss - squared error.

ceml.backend.torch.optimizer

class ceml.backend.torch.optimizer.optimizer.TorchOptimizer(**kwds)

Bases: ceml.optim.optimizer.Optimizer

Wrapper for a PyTorch optimization algorithm.

The TorchOptimizer provides an interface for wrapping an arbitrary PyTorch optimization algorithm (see torch.optim) and minimizing a given loss function.

init(model, loss, x, optim, optim_args, lr_scheduler=None, lr_scheduler_args=None, tol=None, max_iter=1, grad_mask=None, device=torch.device)

Initializes all parameters.

Parameters

  • model (instance of torch.nn.Module) – The model that is to be used.

  • loss (instance of ceml.backend.torch.costfunctions.RegularizedCost) – The loss that has to be minimized.

  • x (numpy.ndarray) – The starting value of x - usually this is the original input whose prediction has to be explained..

  • optim (instance of torch.optim.Optimizer) – Optimizer for minimizing the loss.

  • optim_args (dict) – Arguments of the optimization algorithm (e.g. learning rate, momentum, …)

  • lr_scheduler (Learning rate scheduler (see torch.optim.lr_scheduler)) –

    Learning rate scheduler (see torch.optim.lr_scheduler).

    The default is None.

  • lr_scheduler_args (dict, optional) –

    Arguments of the learning rate scheduler.

    The default is None.

  • tol (float, optional) –

    Tolerance for termination.

    The default is 0.0

  • max_iter (int, optional) –

    Maximum number of iterations.

    The default is 1.

  • grad_mask (numpy.array, optional) –

    Mask that is multiplied element wise on top of the gradient - can be used to hold some dimensions constant.

    If grad_mask is None, no gradient mask is used.

    The default is None.

  • device (torch.device) –

    Specifies the hardware device (e.g. cpu or gpu) we are working on.

    The default is torch.device(“cpu”).

Raises

TypeError – If the type of loss or model is not correct.