corresponding R code

[nick-youngblut]

Any plans on creating corresponding R code to that shown in the README?

[egr95]

Good point. Here's a quick attempt, which assumes your inputs are a row-normalized matrix and a vector of zeros and ones (so it won't work, for example, if y_train is a factor). Note this is tested less thoroughly than the original python code.

aitchison_mixup <- function(X_train, y_train, factor=10, weight=0.5) {  
  X <- X_train
  y <- y_train
  w <- rep(1, length(y))
  
  for (val in unique(y_train)) {
    idxs <- y_train == val
    X_temp <- X_train[idxs, ]
    n <- nrow(X_temp)
    n_aug <- as.integer(factor * n) - n
    
    lam <- runif(n_aug)
    idx1 <- sample(n, n_aug, replace=TRUE)
    idx2 <- sample(n, n_aug, replace=TRUE)
    
    # Take convex combination
    X_aug <- lam * X_temp[idx1, ] + (1 - lam) * X_temp[idx2, ]
    
    X <- rbind(X, X_aug)
    y <- c(y, rep(val, n_aug))
    w <- c(w, rep(weight / (1 - weight) * nrow(X_train) / n_aug, n_aug))
  }
  
  # Shuffle data
  idx <- sample(nrow(X))
  
  return(list(X[idx,], y[idx], w[idx]))
}

compositional_feature_dropout <- function(X_train, y_train, factor=10, weight=0.5) {  
  X <- X_train
  y <- y_train
  w <- rep(1, length(y))
  
  for (val in c(0, 1)) {
    idxs <- y_train == val
    X_temp <- X_train[idxs, ]
    n <- nrow(X_temp)
    n_aug <- as.integer(factor * n) - n
    X_aug <- list()
    y_aug <- list()
    p <- runif(n_aug)
    idx <- sample(n, n_aug, replace=TRUE)
    mask <- matrix(rbinom(n_aug * ncol(X_temp), 1, p), nrow=n_aug, ncol=ncol(X_temp))
    X_new <- X_temp[idx, ]
    X_new[mask==1] <- 1
    X_aug[[1]] <- X_new
    y_aug[[1]] <- y_train[idx]
    X_aug <- do.call(rbind, X_aug)
    y_aug <- unlist(y_aug)
    X <- rbind(X, X_aug)
    y <- c(y, rep(val, n_aug))
    w <- c(w, rep(weight / (1 - weight) * nrow(X_train) / n_aug, n_aug))
  }
  
  # Shuffle data
  idx <- sample(nrow(X))
  
  return(list(X[idx,], y[idx], w[idx]))
}

compositional_cutmix <- function(X_train, y_train, factor=10, weight=0.5) {
  X <- X_train
  y <- y_train
  w <- rep(1, length(y))
  
  for (val in c(0, 1)) {
    idxs <- y_train == val
    X_temp <- X_train[idxs, ]
    n <- nrow(X_temp)
    n_aug <- as.integer(factor * n - n)
    
    idx1 <- sample(n, n_aug, replace=TRUE)
    idx2 <- sample(n, n_aug, replace=TRUE)
    
    p <- runif(n_aug)
    mask <- matrix(rbinom(n_aug * ncol(X_temp), 1, p), nrow=n_aug, ncol=ncol(X_temp))
    X_aug <- mask * X_temp[idx1, ] + (1 - mask) * X_temp[idx2, ]
    
    X_aug <- X_aug / apply(X_aug, 1, sum)
    
    X <- rbind(X, X_aug)
    y <- c(y, rep(val, n_aug))
    w <- c(w, rep(weight / (1 - weight) * nrow(X_train) / n_aug, n_aug))
  }
  
  # Shuffle data
  idx <- sample(nrow(X))
  
  return(list(X[idx,], y[idx], w[idx]))
}