skwdro.base.losses_torch module#

API

Here you can fin the API for all available losses in the library. They all subclass the base Loss class, and so should your custom implementations if you wish to build one from scratch.

In order to make your own loss, you must overload Loss.value() and probably Loss.default_sampler()

class skwdro.base.losses_torch.LogisticLoss(sampler: LabeledSampler, *, d: int = 0, l2reg: float | None = None, fit_intercept: bool = False)[source]

Bases: Loss

Logisic regression with classes \(\{-1, 1\}\)

Parameters:
samplerOptional[LabeledSampler]

sampler for the adversarial samples

dint

dimension of the data (xi.size(-1))

fit_interceptbool

model has an affine dimension

Attributes:
Lnn.SoftMarginLoss

torch solution to implement the soft margin in ]-1, 1[

classifnn.Tanh

activation function to project tensor in ]-1, 1[

linearnn.Linear

linear combination containing the relevant parameters

classmethod default_sampler(xi, xi_labels, epsilon, seed: int | None) → BaseSampler[source]

Overload this method if you wish to define a default sampler associated to your loss.

property intercept: Tensor
predict(X: Tensor) → Tensor[source]

Predict the label of the argument tensor

Parameters:
self

self

Xpt.Tensor

X

Returns:
pt.Tensor
property theta: Tensor
value(xi: Tensor, xi_labels: Tensor) → Tensor[source]
value(xi: Tensor, xi_labels: None) → Tensor

Perform forward pass. Overload the method to implement your own.

class skwdro.base.losses_torch.Loss(sampler: BaseSampler | None, has_labels: bool, *, l2reg: float | None = None, xi: Tensor | None = None, xi_labels: Tensor | None = None, sigma: float | None = None)[source]

Bases: Module, ABC

Base class for loss functions

abstractmethod classmethod default_sampler(xi, xi_labels, epsilon, seed: int | None) → BaseSampler | None[source]

Overload this method if you wish to define a default sampler associated to your loss.

forward(*args)[source]

Alias for value(), for consistency with usual torch api.

has_labels: bool
abstract property intercept
l2reg: Tensor | None
regularize(loss: Tensor) → Tensor[source]

Returns the regularized loss, used in the value function. Adds a new term \(\frac{1}{2}\|\theta\|_2^2\)

sample_pi0(n_samples: int) → Tuple[Tensor, Tensor | None][source]
property sampler: BaseSampler
abstract property theta
value(xi: Tensor, xi_labels: Tensor) → Tensor[source]
value(xi: Tensor, xi_labels: None) → Tensor

Perform forward pass. Overload the method to implement your own.

value_old(theta, xi)[source]

DEPRECATED, DO NOT USE

skwdro.base.losses_torch.NewsVendorLoss

alias of NewsVendorLoss_torch

class skwdro.base.losses_torch.NewsVendorLoss_torch(sampler: NoLabelsSampler, *, k: float = 5, u: float = 7, l2reg: float | None = None, name: str | None = 'NewsVendor loss')[source]

Bases: Loss

Loss associated with the newsvendor problem:

\[k\theta - \mathbb{E}[u\min(\theta, \xi)]\]
Parameters:
samplerOptional[NoLabelsSampler]

optional sampler to use for the demand

kint

journal price

uint

benefit per journal sold

name

name of the loss, optional

Attributes:
samplerNoLabelsSampler
knn.Parameter
unn.Parameter
thetann.Parameter

number of journal stocked

nameOptional[str]
classmethod default_sampler(xi, xi_labels, epsilon, seed: int | None) → NewsVendorNormalSampler[source]

Overload this method if you wish to define a default sampler associated to your loss.

property intercept: None
property theta: Tensor
value(xi: Tensor, xi_labels: Tensor | None = None) → Tensor[source]

Forward pass of the loss on the data

Parameters:
xipt.Tensor

empirical observations of demand

xi_labelsNoneType

placeholder, do not touch

value_old(theta, xi)[source]

DEPRECATED, DO NOT USE

class skwdro.base.losses_torch.QuadraticLoss(sampler: LabeledSampler, *, d: int = 0, l2reg: float | None = None, fit_intercept: bool = False)[source]

Bases: Loss

classmethod default_sampler(xi, xi_labels, epsilon, seed: int | None) → BaseSampler[source]

Overload this method if you wish to define a default sampler associated to your loss.

property intercept: Tensor
regression(X) → Tensor[source]
property theta: Tensor
value(xi: Tensor, xi_labels: Tensor | None) → Tensor[source]

Perform forward pass. Overload the method to implement your own.

class skwdro.base.losses_torch.SimpleWeber(d: int)[source]

Bases: Module

forward(xi: Tensor, xi_labels: Tensor) → Tensor[source]

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

reduction = 'none'
class skwdro.base.losses_torch.WrappedPrimalLoss(loss: Module | Callable[[...], Tensor], transform: Module | None, sampler: BaseSampler, has_labels: bool, reduce_spatial_dims: bool = True, *, l2reg: float | None = None)[source]

Bases: Loss

classmethod default_sampler(xi, xi_labels, epsilon, seed: int | None) → BaseSampler[source]

Overload this method if you wish to define a default sampler associated to your loss.

has_labels: bool
property intercept
loss_oop_interface: bool = True
reduce_spatial_dims: bool = True
property theta
value(xi: Tensor, xi_labels: Tensor | None = None) → Tensor[source]

Perform forward pass. Overload the method to implement your own.