torchgel 0.22.0

PyTorch implementation of group elastic net

torch-gel

This package provides PyTorch implementations to solve the group elastic net problem. Let A_j (j = 1 … p) be feature matrices of sizes m × n_j (m is the number of samples, and n_j is the number of features in the j^th group), and let y be an m × 1 vector of the responses. Group elastic net finds coefficients β_j, and a bias β₀ that solve the optimization problem

min _{β0, …, βp} ½ ║y - β₀ - ∑ A_j β_j║² + m ∑ √n_j (λ₁║β_j║ + λ₂║β_j║²).

Here λ₁ and λ₂ are scalar coefficients that control the amount of 2-norm and squared 2-norm regularization. This 2-norm regularization encourages sparsity at the group level; entire β_j might become 0. The squared 2-norm regularization is in similar spirit to elastic net, and addresses some of the issues of lasso. Note that group elastic net includes as special cases group lasso (λ₂ = 0), ridge regression (λ₁ = 0), elastic net (each n_j = 1), and lasso (each n_j = 1 and λ₂ = 0). The optimization problem is convex, and can be solved efficiently. This package provides two implementations; one based on proximal gradient descent, and one based on coordinate descent.

Installation

Install with pip

pip install torchgel

tqdm (for progress bars) is pulled in as a dependency. PyTorch (v1.0+) is also needed, and needs to be installed manually. Refer to the PyTorch website for instructions.

Solving Single Instances

The modules gel.gelfista and gel.gelcd provide implementations based on proximal gradient descent and coordinate descent respectively. Both have similar interfaces, and expose two main public functions: make_A and gel_solve. The feature matrices should be stored in a list (say As) as PyTorch tensor matrices, and the responses should be stored in a PyTorch vector (say y). Additionally, the sizes of the groups (n_j) should be stored in a vector (say ns). First use the make_A function to convert the feature matrices into a suitable format:

A = make_A(As, ns)

Then pass A, y and other required arguments to gel_solve. The general interface is::

b_0, B = gel_solve(A, y, l_1, l_2, ns, **kwargs)

l_1 and l_2 are floats representing λ₁ and λ₂ respectively. The method returns a float b_0 representing the bias and a PyTorch matrix B holding the other coefficients. B has size p × max_j n_j with suitable zero padding. The following sections cover additional details for the specific implementations.

Proximal Gradient Descent (FISTA)

The gel.gelfista module contains a proximal gradient descent implementation. It's usage is just as described in the template above. Refer to the docstring for gel.gelfista.gel_solve for details about the other arguments.

Coordinate Descent

The gel.gelcd module contains a coordinate descent implementation. Its usage is a bit more involved than the FISTA implementation. Coordinate descent iteratively solves single blocks (each corresponding to a single β_j). There are multiple solvers provided to solve the individual blocks. These are the gel.gelcd.block_solve_* functions. Refer to their docstrings for details about their arguments. gel.gelcd.gel_solve requires passing a block solve function and its arguments (as a dictionary). Refer to its docstring for further details.

Solution Paths

gel.gelpaths provides a wrapper function gel_paths to solve the group elastic net problem for multiple values of the regularization coefficients. It implements a two-stage process. For a given λ₁ and λ₂, first the group elastic net problem is solved and the feature blocks with non-zero coefficients is extracted (the support). Then ridge regression models are learned for each of several provided regularization values. The final model is summarized using an arbitrary provided summary function, and the summary for each combination of the regularization values is returned as a dictionary. The docstring contains more details. gel.ridgepaths contains another useful function, ridge_paths which can efficiently solve ridge regression for multiple regularization values.

Citation

If you find this code useful in your research, please cite

@misc{koushik2017torchgel,
  author = {Koushik, Jayanth},
  title = {torch-gel},
  year = {2017},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/jayanthkoushik/torch-gel}},
}