Reconstruction Criteria

Utility functions required for the GAN based anomaly detection.

class ecgan.utils.reconstruction_criteria.SimilarityCriterions[source]

Bases: abc.ABC

Optimization criterion based on the dissimilarity of samples.

abstract calculate(x, y)[source]

Calculate the (dis-)similarity between two Tensors x and y.

Return type: Tensor

class ecgan.utils.reconstruction_criteria.RBFSimilarityCriterion(**kwargs)[source]

Bases: ecgan.utils.reconstruction_criteria.SimilarityCriterions

Optimization criterion based on the RBF similarity of samples.

calculate(x, y)[source]

Calculate the mean RBF dissimilarity between two tensors of arbitrary shape.

Return type: Tensor

class ecgan.utils.reconstruction_criteria.CosineSimilarityCriterion[source]

Bases: ecgan.utils.reconstruction_criteria.SimilarityCriterions

Optimization criterion based on the cosine similarity of samples.

calculate(x, y)[source]

Calculate the mean cosine dissimilarity between two tensors of arbitrary shape.

Return type: Tensor

class ecgan.utils.reconstruction_criteria.RganMmdCriterion(sigma=None)[source]

Bases: ecgan.utils.reconstruction_criteria.SimilarityCriterions

Optimization criterion based on the MMD similarity of samples.

calculate(x, y)[source]

Pytorch implementation of the RGAN MMD.

The implementation is equivalent to the implementation of the ecgan.utils.reconstruction_criteria._mix_rbf_kernel function() from [RGAN repository, GitHub](https://github.com/ratschlab/RGAN/blob/master/mmd.py). The quadratic-time MMD with Gaussian RBF kernel is computed and - digressing from the original tensorflow implementation - only the K_XY kernel is returned.

ecgan.utils.reconstruction_criteria.rbf_kernel(x, y, rbf_mode=None, sigma=None)[source]

Calculate the Gaussian kernel function between two tensors.

Gaussian kernel between samples of x and y. If sigma is not set, the kernel will infer the sigma value via median pairwise distance.

Parameters

x (Tensor) -- Tensor of shape (N x Features).
y (Tensor) -- Tensor of shape (N x Features).
rbf_mode (Optional[str]) -- Explicit choice of kernel: gaussian, exp or laplacian.
sigma (Optional[float]) -- Sigma for RBF bandwidth. Is usually set automatically by the object but can also be set manually.

Return type

Tensor

Returns

Kernel matrix K(X,Y) with shape [NX, NY]

ecgan.utils.reconstruction_criteria.get_reconstruction_criterion(criterion='residual')[source]

Select criterion function.

Criteria are either distance based or similarity based. The target is usually to reduce either the distance or the dissimilarity (i.e. increase the similarity) between to samples.

Return type: Callable[[Tensor, Tensor], Tensor]
Returns: Callable reconstruction criterion.