bases (version 0.2.0)

b_conv: Random convolutional features

Description

Generates random convolutional features from a list of images. Convolutional kernels are generated randomly (either from a Gaussian distribution or as patches extracted from the training images), applied to each image via efficient matrix multiplication, and then pooled to produce a fixed-size feature vector per image.

Usage

b_conv(
  x,
  p = 100,
  size = 3,
  stride = 1,
  kernel_gen = c("rnorm", "patch"),
  activation = max,
  stdize = c("scale", "box", "symbox", "none"),
  kernels = NULL,
  shift = NULL,
  scale = NULL
)

Value

A matrix of random convolutional features with one row per image in x and one column per kernel (or more columns if activation is multivariate).

Arguments

x

A list of images, where each image is a matrix (for grayscale) or a 3D array with dimensions (height, width, channels) for color images. Images may have different dimensions, but must be large enough to accommodate the convolution kernel size. Missing values are not allowed.

p

The number of random convolutional kernels to generate.

size

The size of the square convolutional kernel (e.g., 3 means a 3x3 kernel).

stride

The stride for the convolution operation, i.e., how many pixels to skip between kernel applications. Default is 1.

kernel_gen

Method for generating convolutional kernels. Either "rnorm" to generate kernels with entries drawn i.i.d. from a standard Normal distribution, or "patch" to extract random patches from the input images.

activation

A function to pool the convolution outputs for each kernel. Defaults to max(). The function should accept a numeric vector and return a scalar or vector of pooled values. Common choices include max(), mean(), functions like the proportion of positive values (PPV), which can be implemented with function(x) mean(x > 0). Multivariate pooling functions are also supported.

stdize

How to standardize the predictors, if at all. The default "scale" applies scale() to the input so that the features have mean zero and unit variance, "box" scales the data along each dimension to lie in the unit hypercube, and "symbox" scales the data along each dimension to lie in \([-0.5, 0.5]^d\).

kernels

Optional matrix of pre-specified convolutional kernels, where each column is a kernel in column-major format. If provided, overrides p, size, and kernel_gen.

shift

Vector of shifts, or single shift value, to use. If provided, overrides those calculated according to stdize.

scale

Vector of scales, or single scale value, to use. If provided, overrides those calculated according to stdize.

Examples

x = outer(1:28, 1:28, function(x, y) {
    d = sqrt(4*(x - 14)^2 + (y - 14)^2)
    dnorm(d, mean = 10, sd = 0.8)
})
pal = gray.colors(256, 1, 0)
image(x, col = pal)

# one random kernel (no activation)
m = b_conv(list(x), p=1, activation=function(x) x)
image(matrix(m, nrow = 26), col = pal)

# many kernels (realistic use case)
m = b_conv(list(x), p = 100, size = 3)
str(m)