Deep learning feature selection

The dl-selection repository contains tools for performing feature selection with deep learning models. It currently has four mechanisms for selecting features, each of which relies on a stochastic relaxation of the feature selection problem. Each mechanism is a learnable input layer that determines which features to select throughout the course of training.

1. Concrete Mask: selects a user-specified number of features k by learning a k-hot vector m for element-wise multiplication with the input x. The layer is composed with a separate network that learns to make predictions using the masked input x * m.

2. Concrete Selector: selects a user-specified number of features k by learning a binary matrix M that selects features from x. The layer is composed with a separate network that learns to make predictions using the selected features Mx.

3. Concrete Gates: selects features subject to a L₀ penalty by learning binary gates m1, m2, ... for each feature. The layer is composed with a separate network that learns to make predictions using the masked input x * m.

4. Concrete Max: selects a user-specified number of features k by learning a Categorical distribution over (1, 2, ..., d) from which features are sampled. The most probable features are selected after training.

Usage

The module selection.models implements a class SelectorMLP for automatically creating a model that composes the user-specified input layer with a prediction network. The model has a built-in train method, so it can be used like this:

import torch.nn as nn
from selection import models

# Load data
train_dataset, val_dataset = ...
input_size, output_size = ...

# Set up model
model = models.SelectorMLP(
    input_layer='concrete_mask',
    k=20,
    input_size=input_size,
    output_size=output_size,
    hidden=[512, 512],
    activation='elu')

# Train model
model.learn(
    train_dataset,
    val_dataset,
    lr=1e-3,
    mbsize=64,
    max_nepochs=300,
    start_temperature=10.0,
    end_temperature=0.01,
    loss_fn=nn.CrossEntropyLoss())

# Extract selected indices
inds = model.get_inds()

Check out the mnist selection.ipynb notebook for examples of how to use each of the layers.

Installation

The easiest way to install this package is with pip:

pip install dl-selection

Or, you can clone the repository to get the most recent version of the code.

Authors

• Ian Covert (icovert@cs.washington.edu)
• Uygar Sümbül
• Su-In Lee