Fix grayscale channel dimension (1) lost in DataLoader

NAMESPACE

@@ -6,6 +6,7 @@ S3method(draw_bounding_boxes,torch_tensor)
 S3method(draw_segmentation_masks,default)
 S3method(draw_segmentation_masks,image_with_segmentation_mask)
 S3method(draw_segmentation_masks,torch_tensor)
+S3method(get_image_size,"magick-image")
 S3method(transform_adjust_brightness,default)
 S3method(transform_adjust_brightness,torch_tensor)
 S3method(transform_adjust_contrast,default)

NEWS.md

@@ -12,6 +12,7 @@
 * Added article showcasing `model_fcn_resnet50()` with visualization utilities `draw_segmentation_masks()` and `vision_make_grid()` (@DerrickUnleashed, #281).
 * Added collection dataset catalog with `search_collection()`, `get_collection_catalog()`, and `list_collection_datasets()` functions for discovering and exploring collections (#271, @ANAMASGARD).
 * Added `target_transform_coco_masks()` and `target_transform_trimap_masks()` transformation functions for explicit segmentation mask generation (@ANAMASGARD).
+* Added support for `connectivity` argument for drawing lines between keypoints in `draw_keypoints()` (@DerrickUnleashed #303)
 
 ## New models
 
@@ -26,11 +27,15 @@
 
 ## Bug fixes and improvements
 
+* `mnist_datataset()` and derivatives now correctly return item x() values with a 1-channel dimension (@Chandraveersingh1717, #307).
+* Fixed `draw_keypoints()` documentation and error message (@srishtiii28, #296).
 * Standardized dataset messages: download messages now include split information, success messages show image count and class count for consistency.
 * fix `model_fasterrcnn_*` did not provide boxes output normalized to image size, did not manage batches, fix performance of the `roi_align()` function (#284)
 * fix rf100 collection bounding-box now consider the correct native COCO format being 'xywh' (#272)
 * Remove `.getbatch` method from MNIST as it is providing inconsistent tensor dimensions with `.getitem` due
 to non-vectorized `transform_` operations (#264)
+* Added article for `draw_keypoints()` (@DerrickUnleashed #303)
+* Fix typos and align model documentation for `model_deeplabv3_*` and `model_convnext_*_detection()` to ensure consistency.(@DerrickUnleashed #302)
 
 # torchvision 0.8.0
 

R/dataset-mnist.R

@@ -118,7 +118,7 @@ mnist_dataset <- dataset(
       archive <- download_and_cache(r[1], prefix = class(self)[1])
       fs::file_copy(archive, destpath)
 
-      if (!tools::md5sum(destpath) == r[2])
+      if (!(tools::md5sum(destpath) == r[2]))
         runtime_error("Corrupt file! Delete the file in {archive} and try again.")
 
     }
@@ -149,18 +149,36 @@ mnist_dataset <- dataset(
   },
 
   .getitem = function(index) {
-    x <- self$data[index, ,]
+    x <- self$data[index, , , drop = FALSE]
     y <- self$targets[index]
 
-    if (!is.null(self$transform))
-      x <- self$transform(x)
+    if (!is.null(self$transform)) {
+        x <- self$transform(x)
+    }
 
-    if (!is.null(self$target_transform))
-      y <- self$target_transform(y)
+    if (!is.null(self$target_transform)) {
+        y <- self$target_transform(y)
+    }
 
     list(x = x, y = y)
   },
 
+  .getbatch = function(index) {
+    x_flat <- self$data[index, , , drop = FALSE]
+    # unsqueeze(2) on array
+    x <- array(x_flat, dim = c(length(index), 1, dim(x_flat)[2:3]))
+    y <- self$targets[index]
+
+    if (!is.null(self$transform)) {
+      x <- torch::torch_stack(lapply(seq_len(dim(x)[1]), function(i) self$transform(x[i,,,])))
+    }
+
+    if (!is.null(self$target_transform)) {
+      y <- self$target_transform(y)
+    }
+
+    list(x = x, y = y)
+  },
   .length = function() {
     dim(self$data)[1]
   },
@@ -179,7 +197,7 @@ mnist_dataset <- dataset(
 #' @describeIn mnist_dataset Kuzushiji-MNIST cursive Japanese character dataset.
 #' @export
 kmnist_dataset <- dataset(
-  name = "kminst_dataset",
+  name = "kmnist_dataset",
   inherit = mnist_dataset,
   archive_size = "21 MB",
   resources = list(
@@ -252,55 +270,63 @@ qmnist_dataset <- dataset(
   },
 
   download = function() {
-    if (self$check_exists())
-      return(NULL)
-
     fs::dir_create(self$raw_folder)
     fs::dir_create(self$processed_folder)
 
-    cli_inform("Downloading {.cls {class(self)[[1]]}} ...")
-    for (r in self$resources[[self$split]]) {
-      filename <- basename(r[1])
-      destpath <- file.path(self$raw_folder, filename)
-
-      archive <- download_and_cache(r[1], prefix = glue::glue("qmnist-{self$split}"))
-      fs::file_copy(archive, destpath, overwrite = TRUE)
-
-      if (!tools::md5sum(destpath) == r[2])
-        runtime_error("Corrupt file! Delete the file in {archive} and try again.")
-    }
-
-    cli_inform("Processing {.cls {class(self)[[1]]}} ...")
-
-
-    if (self$split == "train") {
-      saveRDS(
-        list(
-          read_sn3_pascalvincent(file.path(self$raw_folder, "qmnist-train-images-idx3-ubyte.gz")),
-          read_sn3_pascalvincent(file.path(self$raw_folder, "qmnist-train-labels-idx2-int.gz"))
-        ),
-        file.path(self$processed_folder, self$files$train)
-      )
+    split_exists <- function(split_name) {
+      fs::file_exists(file.path(self$processed_folder, self$files[[split_name]]))
     }
 
-    if (self$split == "test") {
-      saveRDS(
-        list(
-          read_sn3_pascalvincent(file.path(self$raw_folder, "qmnist-test-images-idx3-ubyte.gz")),
-          read_sn3_pascalvincent(file.path(self$raw_folder, "qmnist-test-labels-idx2-int.gz"))
-        ),
-        file.path(self$processed_folder, self$files$test)
-      )
-    }
+    if (all(vapply(names(self$resources), split_exists, logical(1))))
+      return(NULL)
 
-    if (self$split == "nist") {
-      saveRDS(
-        list(
-          read_sn3_pascalvincent(file.path(self$raw_folder, "xnist-images-idx3-ubyte.xz")),
-          read_sn3_pascalvincent(file.path(self$raw_folder, "xnist-labels-idx2-int.xz"))
-        ),
-        file.path(self$processed_folder, self$files$nist)
-      )
+    cli_inform("Downloading {.cls {class(self)[[1]]}} ...")
+    for (split_name in names(self$resources)) {
+      if (split_exists(split_name))
+        next
+
+      for (r in self$resources[[split_name]]) {
+        filename <- basename(r[1])
+        destpath <- file.path(self$raw_folder, filename)
+
+        archive <- download_and_cache(r[1], prefix = glue::glue("qmnist-{split_name}"))
+        fs::file_copy(archive, destpath, overwrite = TRUE)
+
+        if (!(tools::md5sum(destpath) == r[2]))
+          runtime_error("Corrupt file! Delete the file in {archive} and try again.")
+      }
+
+      cli_inform("Processing {.val {split_name}} split for {.cls {class(self)[[1]]}} ...")
+
+      if (split_name == "train") {
+        saveRDS(
+          list(
+            read_sn3_pascalvincent(file.path(self$raw_folder, "qmnist-train-images-idx3-ubyte.gz")),
+            read_sn3_pascalvincent(file.path(self$raw_folder, "qmnist-train-labels-idx2-int.gz"))
+          ),
+          file.path(self$processed_folder, self$files$train)
+        )
+      }
+
+      if (split_name == "test") {
+        saveRDS(
+          list(
+            read_sn3_pascalvincent(file.path(self$raw_folder, "qmnist-test-images-idx3-ubyte.gz")),
+            read_sn3_pascalvincent(file.path(self$raw_folder, "qmnist-test-labels-idx2-int.gz"))
+          ),
+          file.path(self$processed_folder, self$files$test)
+        )
+      }
+
+      if (split_name == "nist") {
+        saveRDS(
+          list(
+            read_sn3_pascalvincent(file.path(self$raw_folder, "xnist-images-idx3-ubyte.xz")),
+            read_sn3_pascalvincent(file.path(self$raw_folder, "xnist-labels-idx2-int.xz"))
+          ),
+          file.path(self$processed_folder, self$files$nist)
+        )
+      }
     }
 
     cli_inform("Dataset {.cls {class(self)[[1]]}} downloaded and extracted successfully.")
@@ -345,6 +371,7 @@ fashion_mnist_dataset <- dataset(
 #' @export
 emnist_collection <- dataset(
   name = "emnist_collection",
+  inherit = mnist_dataset,
   archive_size = "540 MB",
 
   resources = list(
@@ -384,11 +411,10 @@ emnist_collection <- dataset(
     self$split <- match.arg(split, choices = c("train", "test"))
     self$dataset <- match.arg(dataset,  choices = names(self$classes_all_dataset))
     self$root_path <- root
-    self$raw_folder <- file.path(root, class(self)[1], "raw")
-    self$processed_folder <- file.path(root, class(self)[1], "processed")
     self$transform <- transform
     self$target_transform <- target_transform
-    self$class <- self$classes_all_dataset[[self$dataset]]
+    self$classes <- self$classes_all_dataset[[self$dataset]]
+    self$class <- self$classes
 
     if (download) {
       cli_inform("Dataset {.val {self$dataset}} split {.val {self$split}} of {.cls {class(self)[[1]]}} (~{.emph {self$archive_size}}) will be downloaded and processed if not already available.")
@@ -414,7 +440,7 @@ emnist_collection <- dataset(
     url <- self$resources[[1]][1]
     archive <- download_and_cache(url, prefix = class(self)[1])
 
-    if (!tools::md5sum(archive) == self$resources[[1]][2])
+    if (!(tools::md5sum(archive) == self$resources[[1]][2]))
       runtime_error("Corrupt file! Delete the file in {archive} and try again.")
 
     unzip_dir <- file.path(self$raw_folder, "unzipped")
@@ -436,28 +462,23 @@ emnist_collection <- dataset(
     fs::file_exists(file.path(self$processed_folder, self$rds_file(self$split, self$dataset)))
   },
 
-  .getitem = function(index) {
-
-    x <- self$data[index, , ]
-    y <- self$targets[index]
-
-    if (!is.null(self$transform))
-      x <- self$transform(x)
-
-    if (!is.null(self$target_transform))
-      y <- self$target_transform(y)
-
-    list(x = x, y = y)
-  },
-
-  .length = function() {
-    dim(self$data)[1]
-  }
+  active = list(
+    raw_folder = function() {
+      file.path(self$root_path, "emnist_collection", "raw")
+    },
 
+    processed_folder = function() {
+      file.path(self$root_path, "emnist_collection", "processed")
+    }
+  )
 )
 
 read_sn3_pascalvincent <- function(path) {
-  x <- gzfile(path, open = "rb")
+  x <- if (grepl("\\\\.xz$", path, ignore.case = TRUE)) {
+    xzfile(path, open = "rb")
+  } else {
+    gzfile(path, open = "rb")
+  }
   on.exit({close(x)})
 
   magic <- readBin(x, endian = "big", what = integer(), n = 1)

R/transforms-array.R

@@ -5,7 +5,16 @@ transform_to_tensor.array <- function(img) {
   if (length(dim(img)) == 2)
     dim(img) <- c(dim(img), 1)
 
-  res <- torch::torch_tensor(img)$permute(c(3, 1, 2))
+  dims <- dim(img)
+  if (length(dims) != 3)
+    value_error("Expected a 2D or 3D array.")
+
+  # Support both HWC (default image arrays) and CHW (channel-first arrays).
+  if (dims[1] <= 4 && dims[3] > 4) {
+    res <- torch::torch_tensor(img)
+  } else {
+    res <- torch::torch_tensor(img)$permute(c(3, 1, 2))
+  }
 
   if (res$dtype == torch::torch_long())
     res <- res/255

R/transforms-magick.R

@@ -53,6 +53,8 @@
 
 # Utils -------------------------------------------------------------------
 
+#' @method get_image_size magick-image
+#' @export
 `get_image_size.magick-image` <- function(img) {
   info <- magick::image_info(img)
   c(info$width, info$height)