-
-
Notifications
You must be signed in to change notification settings - Fork 8
Allow virtual lazy tensors as targets in classification and regression #386
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Draft
sebffischer
wants to merge
8
commits into
main
Choose a base branch
from
lt-target
base: main
Could not load branches
Branch not found: {{ refName }}
Loading
Could not load tags
Nothing to show
Loading
Are you sure you want to change the base?
Some commits from the old base branch may be removed from the timeline,
and old review comments may become outdated.
Draft
Changes from all commits
Commits
Show all changes
8 commits
Select commit
Hold shift + click to select a range
5447222
...
sebffischer b012de9
...
sebffischer 521fdd7
...
sebffischer a39becc
Merge branch 'main' into lt-target
sebffischer 27cf5b2
Merge branch 'main' into lt-target
sebffischer 9aa26f1
...
sebffischer 06595c7
fixes
sebffischer 61dc13b
...
sebffischer File filter
Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,249 @@ | ||
|
|
||
| #' @title Special Backend for Lazy Tensors | ||
| #' @description | ||
| #' This backend essentially allows you to use a [`torch::dataset`] directly with | ||
| #' an [`mlr3::Learner`]. | ||
| #' | ||
| #' * The data cannot contain missing values, as [`lazy_tensor`]s do not support them. | ||
| #' For this reason, calling `$missings()` will always return `0` for all columns. | ||
| #' * The `$distinct()` method will consider two lazy tensors that refer to the same element of a | ||
| #' [`DataDescriptor`] to be identical. | ||
| #' This means, that it might be underreporting the number of distinct values of lazy tensor columns. | ||
| #' | ||
| #' @export | ||
| #' @examplesIf torch::torch_is_installed() | ||
| #' # used as feature in all backends | ||
| #' x = torch_randn(100, 10) | ||
| #' # regression | ||
| #' ds_regr = tensor_dataset(x = x, y = torch_randn(100, 1)) | ||
| #' be_regr = as_data_backend(ds_regr, converter = list(y = as.numeric)) | ||
| #' be_regr$head() | ||
| #' | ||
| #' | ||
| #' # binary classification: underlying target tensor must be float in [0, 1] | ||
| #' ds_binary = tensor_dataset(x = x, y = torch_randint(0, 2, c(100, 1))$float()) | ||
| #' be_binary = as_data_backend(ds_binary, converter = list( | ||
| #' y = function(x) factor(as.integer(x), levels = c(0, 1), labels = c("A", "yes")) | ||
| #' )) | ||
| #' be_binary$head() | ||
| #' | ||
| #' # multi-class classification: underlying target tensor must be integer in [1, K] | ||
| #' ds_multiclass = tensor_dataset(x = x, y = torch_randint(1, 4, size = c(100, 1))) | ||
| #' be_multiclass = as_data_backend(ds_multiclass, converter = list(y = as.numeric)) | ||
| #' be_multiclass$head() | ||
|
|
||
| DataBackendLazyTensors = R6Class("DataBackendLazyTensors", | ||
| cloneable = FALSE, | ||
| inherit = DataBackendDataTable, | ||
| public = list( | ||
| chunk_size = NULL, | ||
| #' @description | ||
| #' Create a new instance of this [R6][R6::R6Class] class. | ||
| #' @param data (`data.table`)\cr | ||
| #' Data containing (among others) [`lazy_tensor`] columns. | ||
| #' @param primary_key (`character(1)`)\cr | ||
| #' Name of the column used as primary key. | ||
| #' @param converter (named `list()` of `function`s)\cr | ||
| #' A named list of functions that convert the lazy tensor columns to their R representation. | ||
| #' The names must be the names of the columns that need conversion. | ||
| #' @param cache (`character()`)\cr | ||
| #' Names of the columns that should be cached. | ||
| #' Per default, all columns that are converted are cached. | ||
| initialize = function(data, primary_key, converter, cache = names(converter), chunk_size = 100) { | ||
| private$.converter = assert_list(converter, types = "function", any.missing = FALSE) | ||
| assert_subset(names(converter), colnames(data)) | ||
| assert_subset(cache, names(converter), empty.ok = TRUE) | ||
| private$.cached_cols = assert_subset(cache, names(converter)) | ||
| self$chunk_size = assert_int(chunk_size, lower = 1L) | ||
| walk(names(private$.converter), function(nm) { | ||
| if (!inherits(data[[nm]], "lazy_tensor")) { | ||
| stopf("Column '%s' is not a lazy tensor.", nm) | ||
| } | ||
| }) | ||
| super$initialize(data, primary_key) | ||
| # select the column whose name is stored in primary_key from private$.data but keep its name | ||
| private$.data_cache = private$.data[, primary_key, with = FALSE] | ||
| }, | ||
| data = function(rows, cols) { | ||
| rows = assert_integerish(rows, coerce = TRUE) | ||
| assert_names(cols, type = "unique") | ||
|
|
||
| if (getOption("mlr3torch.data_loading", FALSE)) { | ||
| # no caching, no materialization as this is called in the training loop | ||
| return(super$data(rows, cols)) | ||
| } | ||
| if (all(intersect(cols, private$.cached_cols) %in% names(private$.data_cache))) { | ||
| expensive_cols = intersect(cols, private$.cached_cols) | ||
| other_cols = setdiff(cols, expensive_cols) | ||
| cache_hit = private$.data_cache[list(rows), expensive_cols, on = self$primary_key, with = FALSE] | ||
| complete = complete.cases(cache_hit) | ||
| cache_hit = cache_hit[complete] | ||
| if (nrow(cache_hit) == length(rows)) { | ||
| tbl = cbind(cache_hit, super$data(rows, other_cols)) | ||
| setcolorder(tbl, cols) | ||
| return(tbl) | ||
| } | ||
| combined = rbindlist(list(cache_hit, private$.load_and_cache(rows[!complete], expensive_cols))) | ||
| reorder = vector("integer", nrow(combined)) | ||
| reorder[complete] = seq_len(nrow(cache_hit)) | ||
| reorder[!complete] = nrow(cache_hit) + seq_len(nrow(combined) - nrow(cache_hit)) | ||
|
|
||
| tbl = cbind(combined[reorder], super$data(rows, other_cols)) | ||
| setcolorder(tbl, cols) | ||
| return(tbl) | ||
| } | ||
|
|
||
| private$.load_and_cache(rows, cols) | ||
| }, | ||
| head = function(n = 6L) { | ||
| if (getOption("mlr3torch.data_loading", FALSE)) { | ||
| return(super$head(n)) | ||
| } | ||
|
|
||
| self$data(seq_len(n), self$colnames) | ||
| }, | ||
| missings = function(rows, cols) { | ||
| set_names(rep(0L, length(cols)), cols) | ||
| } | ||
| ), | ||
| active = list( | ||
| converter = function(rhs) { | ||
| assert_ro_binding(rhs) | ||
| private$.converter | ||
| } | ||
| ), | ||
| private = list( | ||
| # call this function only with rows that are not in the cache yet | ||
| .load_and_cache = function(rows, cols) { | ||
| # Process columns that need conversion | ||
| tbl = super$data(rows, cols) | ||
| cols_to_convert = intersect(names(private$.converter), names(tbl)) | ||
| tbl_to_mat = tbl[, cols_to_convert, with = FALSE] | ||
| # chunk the rows of tbl_to_mat into chunks of size self$chunk_size, apply materialize | ||
| n = nrow(tbl_to_mat) | ||
| chunks = split(seq_len(n), rep(seq_len(ceiling(n / self$chunk_size)), each = self$chunk_size, length.out = n)) | ||
|
|
||
| tbl_mat = if (n == 0) { | ||
| set_names(list(torch_empty(0)), names(tbl_to_mat)) | ||
| } else { | ||
| set_names(lapply(transpose_list(lapply(chunks, function(chunk) { | ||
| materialize(tbl_to_mat[chunk, ], rbind = TRUE) | ||
| })), torch_cat, dim = 1L), names(tbl_to_mat)) | ||
| } | ||
|
|
||
| for (nm in cols_to_convert) { | ||
| converted = private$.converter[[nm]](tbl_mat[[nm]]) | ||
| tbl[[nm]] = converted | ||
|
|
||
| if (nm %in% private$.cached_cols) { | ||
| set(private$.data_cache, i = rows, j = nm, value = converted) | ||
| } | ||
| } | ||
| return(tbl) | ||
| }, | ||
| .data_cache = NULL, | ||
| .converter = NULL, | ||
| .cached_cols = NULL | ||
| ) | ||
| ) | ||
|
|
||
| #' @export | ||
| as_data_backend.dataset = function(x, dataset_shapes, ...) { | ||
| tbl = as_lazy_tensors(x, dataset_shapes, ...) | ||
| tbl$row_id = seq_len(nrow(tbl)) | ||
| DataBackendLazyTensors$new(tbl, primary_key = "row_id", ...) | ||
| } | ||
|
|
||
| #' @export | ||
| as_task_classif.dataset = function(x, target, levels, converter = NULL, dataset_shapes = NULL, chunk_size = 100, cache = names(converter), ...) { | ||
| if (length(x) < 2) { | ||
| stopf("Dataset must have at least 2 rows.") | ||
| } | ||
| batch = dataloader(x, batch_size = 2)$.iter()$.next() | ||
| if (is.null(converter)) { | ||
| if (length(levels) == 2) { | ||
| if (batch[[target]]$dtype != torch_float()) { | ||
| stopf("Target must be a float tensor, but has dtype %s", batch[[target]]$dtype) | ||
| } | ||
| if (test_equal(batch[[target]]$shape, c(2L, 1L))) { | ||
| converter = set_names(list(crate(function(x) factor(as.integer(x), levels = 0:1, labels = levels), levels)), target) | ||
| } else { | ||
| stopf("Target must be a float tensor of shape (batch_size, 1), but has shape (batch_size, %s)", | ||
| paste(batch[[target]]$shape[-1L], collapse = ", ")) | ||
| } | ||
| converter = set_names(list(crate(function(x) factor(as.integer(x), levels = 0:1, labels = levels), levels)), target) | ||
| } else { | ||
| if (batch[[target]]$dtype != torch_int()) { | ||
| stopf("Target must be an integer tensor, but has dtype %s", batch[[target]]$dtype) | ||
| } | ||
| if (test_equal(batch[[target]]$shape, 2L)) { | ||
| converter = set_names(list(crate(function(x) factor(as.integer(x), labels = levels), levels)), target) | ||
| } else { | ||
| stopf("Target must be an integer tensor of shape (batch_size), but has shape (batch_size, %s)", | ||
| paste(batch[[target]]$shape[-1L], collapse = ", ")) | ||
| } | ||
| converter = set_names(list(crate(function(x) factor(as.integer(x), labels = levels), levels)), target) | ||
| } | ||
| } | ||
| be = as_data_backend(x, dataset_shapes, converter = converter, cache = cache, chunk_size = chunk_size) | ||
| as_task_classif(be, target = target, ...) | ||
| } | ||
|
|
||
| #' @export | ||
| as_task_regr.dataset = function(x, target, converter = NULL, dataset_shapes = NULL, chunk_size = 100, cache = names(converter), ...) { | ||
| if (length(x) < 2) { | ||
| stopf("Dataset must have at least 2 rows.") | ||
| } | ||
| if (is.null(converter)) { | ||
| converter = set_names(list(as.numeric), target) | ||
| } | ||
| batch = dataloader(x, batch_size = 2)$.iter()$.next() | ||
|
|
||
| if (batch[[target]]$dtype != torch_float()) { | ||
| stopf("Target must be a float tensor, but has dtype %s", batch[[target]]$dtype) | ||
| } | ||
|
|
||
| if (!test_equal(batch[[target]]$shape, c(2L, 1L))) { | ||
| stopf("Target must be a float tensor of shape (batch_size, 1), but has shape (batch_size, %s)", | ||
| paste(batch[[target]]$shape[-1L], collapse = ", ")) | ||
| } | ||
|
|
||
| dataset_shapes = get_or_check_dataset_shapes(x, dataset_shapes) | ||
| be = as_data_backend(x, dataset_shapes, converter = converter, cache = cache, chunk_size = chunk_size) | ||
| as_task_regr(be, target = target, ...) | ||
| } | ||
|
|
||
| #' @export | ||
| col_info.DataBackendLazyTensors = function(x, ...) { # nolint | ||
| first_row = x$head(1L) | ||
| types = map_chr(first_row, function(x) class(x)[1L]) | ||
| discrete = setdiff(names(types)[types %chin% c("factor", "ordered")], x$primary_key) | ||
| levels = insert_named(named_list(names(types)), map(first_row[, discrete, with = FALSE], levels)) | ||
| data.table(id = names(types), type = unname(types), levels = levels, key = "id") | ||
| } | ||
|
|
||
|
|
||
| # conservative check that avoids that a pseudo-lazy-tensor is preprocessed by some pipeop | ||
| # @param be | ||
| # the backend | ||
| # @param candidates | ||
| # the feature and target names | ||
| # @param visited | ||
| # Union of all colnames already visited | ||
| # @return visited | ||
| check_lazy_tensors_backend = function(be, candidates, visited = character()) { | ||
| if (inherits(be, "DataBackendRbind") || inherits(be, "DataBackendCbind")) { | ||
| bs = be$.__enclos_env__$private$.data | ||
| # first we check b2, then b1, because b2 possibly overshadows some b1 rows/cols | ||
| visited = check_lazy_tensors_backend(bs$b2, candidates, visited) | ||
| check_lazy_tensors_backend(bs$b1, candidates, visited) | ||
| } else { | ||
| if (inherits(be, "DataBackendLazyTensors")) { | ||
| if (any(names(be$converter) %in% visited)) { | ||
| converter_cols = names(be$converter)[names(be$converter) %in% visited] | ||
| stopf("A converter column ('%s') from a DataBackendLazyTensors was presumably preprocessed by some PipeOp. This can cause inefficiencies and is therefore not allowed. If you want to preprocess them, please directly encode them as R types.", paste0(converter_cols, collapse = ", ")) # nolint | ||
| } | ||
| } | ||
| union(visited, intersect(candidates, be$colnames)) | ||
| } | ||
| } |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
since I have not used lazy tensors very much, it would help to see a more detailed description of changes here in NEWS. I also would expect some changes to https://mlr3torch.mlr-org.com/articles/lazy_tensor.html but I do not see any in this PR yet.
What is the typical use case which motivates this PR?