R/cityscapes-pix2pix.R

In torchdatasets: Ready to Use Extra Datasets for Torch

#' Cityscapes Pix2Pix dataset
#'
#' Downloads and prepares the cityscapes dataset that has been used in the
#' [pix2pix paper](https://arxiv.org/abs/1611.07004).
#'
#' Find more information in the [project website](https://phillipi.github.io/pix2pix/)
#'
#' @inheritParams bird_species_dataset
#' @inheritParams torchvision::image_folder_dataset
#' @param ... Currently unused.
#'
#' @export
cityscapes_pix2pix_dataset <- torch::dataset(
  "CityscapesImagePairs",
  initialize = function(root, split = "train", download = FALSE, ...,
                        transform = NULL, target_transform = NULL) {

    url <- "http://efrosgans.eecs.berkeley.edu/pix2pix/datasets/cityscapes.tar.gz"

    data_path <- maybe_download(
      root = root,
      url = url,
      name = "cityscapes-image-pairs",
      download = download,
      extract_fun = function(f, exdir) {
        untar(f, exdir = exdir)
      }
    )

    self$split <- split

    path <- fs::path(
      data_path,
      "cityscapes",
      ifelse(self$split == "train", "train", "val")
    )

    self$files <- fs::dir_ls(path, glob = "*.jpg")
    self$transform <- if (is.null(transform)) identity else transform
    self$target_transform <- if (is.null(target_transform)) identity else target_transform
  },
  .getitem = function(i) {
    img <- jpeg::readJPEG(self$files[i])

    list(
      input_img = self$transform(img[,257:512,]),
      real_img = self$target_transform(img[,1:256,])
    )
  },
  .length = function() {
    length(self$files)
  }
)