Add VGGFace2 Dataset #238
Add VGGFace2 Dataset #238
Conversation
|
@cregouby This is what has been implemented ds <- vggface2_dataset(download = TRUE)
item <- ds[1]
item$x # image array
item$y # integer label
ds$classes[item$y] # list(name=..., gender=...)We can also implement (not yet implemented) a method to give more details, like to the ds$classes but one major problem would be this data is available only for select pictures so it would have How can we proceed from here ? |
|
Sure @cregouby I'll make the necessary changes. I had exams so wasn't available for sometime. Will work on it asap. |
cregouby
left a comment
cregouby
left a comment
There was a problem hiding this comment.
praise thanks for switching to clear and readable resourse definition.
todo The dataset is for instance segmentation task and as such requires an instance segmentation output, not a classification output.
|
|
||
| cli_inform("Downloading {.cls {class(self)[[1]]}}...") | ||
|
|
||
| for (i in seq_len(nrow(self$resources))) { |
There was a problem hiding this comment.
question Do we need to download 36GB of data (train-set) when the end-user request train=FALSE ?
suggestion I would save user time and disk space by limiting the download to the requested split
|
|
||
| for (i in seq_len(nrow(self$resources))) { | ||
| row <- self$resources[i, ] | ||
| archive <- download_and_cache(row$url, prefix = row$split) |
There was a problem hiding this comment.
todo could we prepend class(self)[[1]] to row$split for prefix= ? (this is to avoid 5 different files, hard to identify as being part of vggface2, spread in the root cache folder)
| #' ds$classes[item$y] # list(name=..., gender=...) | ||
| #' } | ||
| #' | ||
| #' @family segmentation_dataset |
There was a problem hiding this comment.
question this is (the first) instance segmentation dataset in the repo. Shall we mix it with object segmentation datasets @family segmentation_dataset or shall we define a new family ? Defining a new family would require an update of the website (see dataset categories in the _pkgdown.yml file) and a test/update for proper management by downstream functions (draw_segmentation_mask, ...)
suggestion let it like this but create an issue "vggface2 is an instance segmentation dataset" with a todo list on that.
| expect_length(vgg, 169396) | ||
| first_item <- vgg[1] | ||
| expect_named(first_item, c("x", "y")) | ||
| expect_type(first_item$x, "double") |
There was a problem hiding this comment.
todo missing as you know the first item, it would be nice to add a check on dim() of the first_item$x object. This gives a hit of image size to folks.
| #' | ||
| #' @return A torch dataset object `vggface2_dataset`: | ||
| #' - `x`: RGB image array. | ||
| #' - `y`: Integer label (1…N) for the identity. |
There was a problem hiding this comment.
todo blocking y integer label is the output of a classification dataset, not an instance segmentation dataset. Either change one or the other
There was a problem hiding this comment.
todo missing as there may be impact on draw_segmentation_mask could we add a check in one of the tests for drawing segmentation mask of the item ?
| if (!is.null(self$target_transform)) { | ||
| y <- self$target_transform(y) | ||
| } | ||
| list(x = x, y = y) |
There was a problem hiding this comment.
todo please add the segmentation class to the output
2 participants
This PR Adds support to load the VGGFace2 Dataset
It also included proper documentation and tests for the dataset
Closes #224