sparse.nn.model.LISTA

class sparse.nn.model.LISTA(in_features, out_features, n_folds=2, solver=BasisPursuitADMM(lambd=0.1, tol=0.0001, max_iters=100))[source]

Bases: Module

Learned Iterative Shrinkage-Thresholding Algorithm [1] AutoEncoder neural network for sparse coding.

Parameters
in_featuresint

The num. of input features (X dimension).

out_featuresint

The dimensionality of the embedding vector Z.

n_foldsint

The num. of recursions to apply to get better convergence of Z. Must be greater or equal to 1.

solverBasisPursuitADMM

Matching Pursuit solver for the \(Q_1^\epsilon\) problem (see sparse.nn.solver.basis_pursuit_admm()). Used only in forward_best function.

Notes

In overcomplete coding, where sparse representations emerge, out_features >> in_features. If out_features ≲ in_features, the encoding representation will be dense.

References

1

Gregor, K., & LeCun, Y. (2010, June). Learning fast approximations of sparse coding. In Proceedings of the 27th international conference on international conference on machine learning (pp. 399-406).

Methods

forward(x)

AutoEncoder forward pass.

forward_best(x)

Test function to match the output of the MatchingPursuit.

Attributes

lambd

Learned Softshrink threshold vector of size out_features.
forward(x)[source]

AutoEncoder forward pass.

Parameters
x(B, C, H, W) torch.Tensor

A batch of input images

Returns
z(B, Z) torch.Tensor

Embedding vectors: sparse representation of x.

decoded(B, C, H, W) torch.Tensor

Reconstructed x from z.

forward_best(x)[source]

Test function to match the output of the MatchingPursuit.

Parameters
x(B, C, H, W) torch.Tensor

A batch of input images

Returns
z(B, Z) torch.Tensor

Embedding vectors: sparse representation of x.

decoded(B, C, H, W) torch.Tensor

Reconstructed x from z.

property lambd

Learned Softshrink threshold vector of size out_features.

reset_parameters()[source]