Feat: Add SAHI (Slicing Aided Hyper Inference) Transform

[DerrickUnleashed]

This PR Adds:

• The implementation of transform_sahi_crop()
• The implementation of target_transform_sahi_crop()
• Adds test suites for the same
• Adds export command to fix this error

✖ transforms-tensor.R:480: S3 method
  `get_image_size.torch_tensor` needs
  @export or @exportS3Method tag.

• Updates nms.Rd Documentation to Roxygen2
  Fixes Implement SAHI: (Slicing Aided Hyper Inference) as both transform_ and target_transform_ #324

[DerrickUnleashed]

@cregouby, I'm still a bit unclear about the usage of target_transform_sahi_crop() However, the transform_sahi_crop() seems to be working as expected

[cregouby]

praise I think that the cropping values logic is there.
todo design all transform_ functions shall return a object of the same type as the input (except transform to tensor) as they will be piped. Your only freedom is to change the shape of the tensor, and move the crops into a sequential batch, as transform_five_crop() do.
ex for x$shape [1, 3, 10, 10], transform_sahi_crop(x, c(4,4) , c(.2, .2)) shall output a tensor of shape [25, 3, 3, 3] (see

torchvision/tests/testthat/test-transforms-tensor.R

Lines 161 to 174 in ac32fa5

	test_that("five_crop", {
	x <- torch_randn(3, 10, 12)
	o <- transform_five_crop(x, c(3, 3))
	expect_length(o, 5)
	expect_tensor_shape(o[[1]], c(3,3,3))
	ob <- transform_five_crop(x$unsqueeze(1), c(3, 3))
	expect_length(ob, 5)
	expect_tensor_shape(ob[[1]], c(1,3,3,3))
	})

)
suggestion you should rely on the existing crop function to crop the tensor, saving you a lot of unit tests.
todo software design this transform shall be an S3 transform so shall have an entry in transform-generics.R for the dispatch, one in transforms-defaults.R, one in transforms-magick.R, one in transforms-tensor.R, as you never know what transform goes before, what transform is piped after.

todo code relocation Please keep (the badly named, my fault) transform-segmentation.R file for target_transforms
suggestion for clarity, you may rename the transform-segmentation.R into target-transform-segmentation.R (and the test file accordingly)

thought as SAHI intimately rely on input image size and coco bbox (so both x and y at the same time), and as there is no way to pass information from transform_sahi_crop to target_transform_sahi_crop, I think maybe a prepare_sahi_crop(dataset, size, overlap_size_ratio, ... ) function gathering all the needed context into a specific class object sahi_preparation and having both functions using it a the only input parameter transform_sahi_crop(x, sahi_preparation) and target_transform_sahi_crop(x, sahi_preparation) could be a nice API to SAHI.

[DerrickUnleashed]

todo software design this transform shall be an S3 transform so shall have an entry in transform-generics.R for the dispatch, one in transforms-defaults.R, one in transforms-magick.R, one in transforms-tensor.R, as you never know what transform goes before, what transform is piped after.

I didn't understand properly, Do I need to move the code of transform_sahi and target_transform_sahi to transform-generics ?

[cregouby]

No. Transform_sahi_crop must manage magick images, tensors and nothing else. The way to do it is S3 dispatch which required an entry in each of the mentioned R files. Le 13 juin 2026 19:56:59 GMT+02:00, Derrick Richard ***@***.***> a écrit :

…

DerrickUnleashed left a comment (mlverse/torchvision#329) > **todo** **software design** this transform shall be an S3 transform so shall have an entry in transform-generics.R for the dispatch, one in transforms-defaults.R, one in transforms-magick.R, one in transforms-tensor.R, as you never know what transform goes before, what transform is piped after. I didn't understand properly, Do I need to move the code of transform_sahi and target_transform_sahi to transform-generics ? -- Reply to this email directly or view it on GitHub: #329 (comment) You are receiving this because you were mentioned. Message ID: ***@***.***>

[DerrickUnleashed]

# From a dataset (SAHI as part of the transform pipeline)
ds <- coco_detection_dataset(train = FALSE, year = "2017", download = TRUE)
sp_ds <- prepare_sahi_split(ds, size = c(200, 200), overlap_size_ratio = c(0.2, 0.2))

ds <- coco_detection_dataset(train = FALSE, year = "2017",
  transform = . %>% transform_to_tensor() %>%
    transform_sahi_crop(sp_ds))

item <- ds[1]
grid <- vision_make_grid(item$x, scale = TRUE, num_rows = 3)
tensor_image_browse(grid)

img_url <- "https://raw.githubusercontent.com/obss/sahi/main/demo/demo_data/small-vehicles1.jpeg"
img <- base_loader(img_url) %>% transform_to_tensor()

sp <- prepare_sahi_split(img, size = c(512, 512))

crops <- transform_sahi_crop(img, sp)

# Synthetic target with a box straddling the first two crops
y <- list(
  boxes = torch_tensor(matrix(c(400, 100, 600, 300), nrow = 1, byrow = TRUE),
                       dtype = torch_float()),
  labels = "car"
)

targets <- target_transform_sahi_crop(y, sp, min_area_ratio = 0.1)

targets[[1]]$boxes  # Box clipped and translated to first crop coordinates

targets[[2]]$boxes  # Box in second crop (the portion that spilled over)

# Visualize the crops with bounding boxes overlaid
preview <- lapply(1:dim(crops)[1], function(i) {
  item <- list(x = crops[i, ..], y = targets[[i]])
  class(item) <- "image_with_bounding_box"
  draw_bounding_boxes(item, colors = "red")
})
grid <- vision_make_grid(torch_stack(preview), scale = FALSE, num_rows = 3)
tensor_image_browse(grid)

[DerrickUnleashed]

thought the question now is how many input S3 format shall prepare_sahi_split() cover ? I'd start with the usual magick image then tensor image (single and batch of) and also cover the dataset..

it supports magick and tensor images as input now

[cregouby]

praise very significant improvement toward the merge
todo see inline

[cregouby]

.

[DerrickUnleashed]

## Not run: 
# Full SAHI pipeline: prepare split, crop image, adjust targets
img_url <- "https://raw.githubusercontent.com/obss/sahi/main/demo/demo_data/small-vehicles1.jpeg"
img <- base_loader(img_url) %>% transform_to_tensor()

sp <- prepare_sahi_split(img, size = c(512, 512), overlap_size_ratio = c(0.2, 0.2))

crops <- transform_sahi_crop(img, sp)
crops$shape

# Synthetic target with a box straddling the first two crops
y <- list(
  boxes = torch_tensor(matrix(c(400, 100, 600, 300), nrow = 1, byrow = TRUE),
                       dtype = torch_float()),
  labels = "car"
)
targets <- target_transform_sahi_crop(y, sp, min_area_ratio = 0.1)

targets[[1]]$boxes  # Box clipped and translated to first crop
targets[[2]]$boxes  # Box in second crop (the portion that spilled over)

# Visualize crops with bounding boxes overlaid
preview <- lapply(1:dim(crops)[1], function(i) {
  item <- list(x = crops[i, ..], y = targets[[i]])
  class(item) <- "image_with_bounding_box"
  draw_bounding_boxes(item, colors = "red")
})
grid <- vision_make_grid(torch_stack(preview), scale = FALSE, num_rows = 3)
tensor_image_browse(grid)

## End(Not run)

[cregouby]

praise thanks for your modifications.
todo a small effort again for completness.