Item transform affine - #367
Conversation
| orig_h <- as.numeric(x$x$shape[length(x$x$shape) - 1]) | ||
| orig_w <- as.numeric(x$x$shape[length(x$x$shape)]) |
There was a problem hiding this comment.
todo readability
| orig_h <- as.numeric(x$x$shape[length(x$x$shape) - 1]) | |
| orig_w <- as.numeric(x$x$shape[length(x$x$shape)]) | |
| c(orig_h, orig_w) %<-% tail(x$x$shape, 2) |
| x | ||
| } | ||
|
|
||
| affine_boxes_xyxy <- function(boxes, angle, translate, scale, shear, center, |
There was a problem hiding this comment.
todo Please rename the function into target_transform_affine.list() and move into target_transform_detection.R
todo Please document and export with the #' @family target_transforms_detection
todo Please add a generic method target_transform_affine() and a default method for proper S3 dispatch
| make_seg_item <- function(image_size = c(100L, 200L), num_masks = 3L) { | ||
| x <- torch_randn(3, image_size[1], image_size[2]) | ||
| masks <- torch_rand(num_masks, image_size[1], image_size[2]) > 0.5 | ||
| labels <- torch_ones(num_masks, dtype = torch_long()) | ||
| y <- list( | ||
| masks = masks, | ||
| labels = labels, | ||
| image_height = image_size[1], | ||
| image_width = image_size[2] | ||
| ) | ||
| item <- list(x = x, y = y) | ||
| class(item) <- c("image_with_segmentation_mask", "list") | ||
| item | ||
| } |
There was a problem hiding this comment.
todo please move this helper function into helper-torchvision.R
todo please rename it into make_segmentation_item for more transparency
There was a problem hiding this comment.
todo Please check the helper function may already be present according to #366
todo Please add a test showing affine can be composed like in
torchvision/tests/testthat/test-item-transforms-geometry.R
Lines 191 to 202 in 384aa14
todo Please note that
make_item() is now make_detection_item()
f6437af to
6904e69
Compare
|
todo the current target transform is not correct, it should be rotated, translated and scaled, around the given center, as the image is. on the provided example and should be fixed : |
cregouby
left a comment
cregouby
left a comment
There was a problem hiding this comment.
Praise This is a major one ! Thanks
| masks <- x$y$masks | ||
| if (!is.null(masks) && masks$ndim >= 3) { | ||
| dtype <- masks$dtype | ||
| masks <- transform_affine(masks, angle = angle, translate = translate, | ||
| scale = scale, shear = shear, | ||
| interpolation = 0, fill = fill, center = center) | ||
| x$y$masks <- masks$to(dtype = dtype) | ||
| } | ||
| x | ||
| } |
There was a problem hiding this comment.
todo Can we move that target_transform_affine.list() as well ?
thought I know this will require ugly test on the presence of parameters, waiting for #364
There was a problem hiding this comment.
todo missing can we manage also the .image_with_rotated_box() S3 class ?
6904e69 to
3830e65
Compare
cregouby
left a comment
cregouby
left a comment
There was a problem hiding this comment.
todo can you resolve the merge conflicts, please ?
3830e65 to
bad7f90
Compare
2 participants
Closes #347.
Adds
item_transform_affine(), applying an affine transform that is rotation, translation, scale and shear to a dataset item and keeping the target aligned with the image. Follows the S3-dispatch convention from #343.