feat: lazy tensor datatype

.github/dependabot.yaml

@@ -0,0 +1,6 @@
+version: 2
+updates:
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"

.github/workflows/pkgdown.yml

@@ -22,6 +22,7 @@ jobs:
       group: pkgdown-${{ github.event_name != 'pull_request' || github.run_id }}
     env:
       GITHUB_PAT: ${{ secrets.GITHUB_TOKEN }}
+      TORCH_INSTALL: 1
     steps:
       - uses: actions/checkout@v3
 
@@ -34,10 +35,6 @@ jobs:
           extra-packages: any::pkgdown, local::.
           needs: website
 
-      - name: Install torch
-        run: torch::install_torch()
-        shell: Rscript {0}
-
       - name: Install template
         run: pak::pkg_install("mlr-org/mlr3pkgdowntemplate")
         shell: Rscript {0}

.github/workflows/r-cmd-check.yml

@@ -28,13 +28,16 @@ jobs:
 
     env:
       GITHUB_PAT: ${{ secrets.GITHUB_TOKEN }}
+      TORCH_INSTALL: 1
 
     strategy:
       fail-fast: false
       matrix:
         config:
           - {os: ubuntu-latest,   r: 'devel'}
           - {os: ubuntu-latest,   r: 'release'}
+          - {os: macos-latest,   r: 'release'}
+          - {os: windows-latest, r: 'release'}
 
     steps:
       - uses: actions/checkout@v3
@@ -50,10 +53,6 @@ jobs:
           extra-packages: any::rcmdcheck
           needs: check
 
-      - name: Install torch
-        run: torch::install_torch()
-        shell: Rscript {0}
-
       - uses: r-lib/actions/check-r-package@v2
 
       - uses: mxschmitt/action-tmate@v3

.lintr

@@ -1,9 +1,9 @@
-linters: with_defaults(
-    # lintr defaults: https://github.com/jimhester/lintr#available-linters
-    # the following setup changes/removes certain linters
-    assignment_linter = NULL, # do not force using <- for assignments
-    object_name_linter = object_name_linter(c("snake_case", "CamelCase")), # only allow snake case and camel case object names
-    cyclocomp_linter = NULL, # do not check function complexity
-    commented_code_linter = NULL, # allow code in comments
-    line_length_linter = line_length_linter(180)
-    )
+linters: linters_with_defaults(
+    assignment_linter=NULL,
+    object_name_linter=NULL,
+    cyclocomp_linter=NULL,
+    commented_code_linter=NULL,
+    line_length_linter=line_length_linter(180),
+    indentation_linter=NULL,
+    object_length_linter=object_length_linter(40))
+

DESCRIPTION

@@ -42,7 +42,7 @@ Imports:
     checkmate (>= 2.2.0),
     coro,
     lgr,
-    mlr3misc (>= 0.11.0),
+    mlr3misc (>= 0.14.0),
     methods,
     data.table,
     paradox (>= 0.11.0),
@@ -59,19 +59,19 @@ Suggests:
     progress,
     rmarkdown,
     viridis,
-    torchvision,
     testthat (>= 3.0.0),
-    zip
+    torchvision
 Remotes:
-    r-lib/zip,
-    mlr-org/mlr3
+    mlr-org/mlr3,
+    mlr-org/mlr3pipelines@feat/keep_results,
+    mlverse/torchvision
 Config/testthat/edition: 3
 NeedsCompilation: no
 ByteCompile: no
 VignetteBuilder: knitr
 Encoding: UTF-8
 Roxygen: list(markdown = TRUE, r6 = TRUE)
-RoxygenNote: 7.2.3.9000
+RoxygenNote: 7.3.1
 Collate:
     'CallbackSet.R'
     'zzz.R'
@@ -81,6 +81,8 @@ Collate:
     'CallbackSetProgress.R'
     'ContextTorch.R'
     'DataBackendLazy.R'
+    'utils.R'
+    'DataDescriptor.R'
     'LearnerTorch.R'
     'LearnerTorchImage.R'
     'LearnerTorchAlexNet.R'
@@ -90,6 +92,7 @@ Collate:
     'ModelDescriptor.R'
     'PipeOpModule.R'
     'PipeOpTorch.R'
+    'PipeOpTaskPreprocTorch.R'
     'PipeOpTorchActivation.R'
     'PipeOpTorchAvgPool.R'
     'PipeOpTorchBatchNorm.R'
@@ -98,6 +101,7 @@ Collate:
     'PipeOpTorchConvTranspose.R'
     'PipeOpTorchDropout.R'
     'PipeOpTorchHead.R'
+    'shape.R'
     'PipeOpTorchIngress.R'
     'PipeOpTorchLayerNorm.R'
     'PipeOpTorchLinear.R'
@@ -110,17 +114,21 @@ Collate:
     'PipeOpTorchReshape.R'
     'PipeOpTorchSoftmax.R'
     'ResamplingRowRoles.R'
+    'TaskClassif_lazy_iris.R'
+    'TaskClassif_mnist.R'
     'TaskClassif_tiny_imagenet.R'
     'TorchDescriptor.R'
     'TorchOptimizer.R'
     'bibentries.R'
     'cache.R'
-    'imageuri.R'
+    'lazy_tensor.R'
     'learner_torch_methods.R'
+    'materialize.R'
+    'merge_graphs.R'
     'nn_graph.R'
     'paramset_torchlearner.R'
+    'preprocess.R'
     'rd_info.R'
     'reset_last_layer.R'
     'task_dataset.R'
-    'utils.R'
     'with_torch_settings.R'

NAMESPACE

@@ -1,14 +1,17 @@
 # Generated by roxygen2: do not edit by hand
 
-S3method("[",imageuri)
-S3method("[<-",imageuri)
-S3method("[[",imageuri)
-S3method("[[<-",imageuri)
+S3method("==",lazy_tensor)
+S3method("[",lazy_tensor)
+S3method("[<-",lazy_tensor)
+S3method("[[",lazy_tensor)
+S3method("[[<-",lazy_tensor)
 S3method(as.data.table,DictionaryMlr3torchCallbacks)
 S3method(as.data.table,DictionaryMlr3torchLosses)
 S3method(as.data.table,DictionaryMlr3torchOptimizers)
-S3method(as_imageuri,character)
-S3method(as_imageuri,imageuri)
+S3method(as_lazy_tensor,DataDescriptor)
+S3method(as_lazy_tensor,dataset)
+S3method(as_lazy_tensor,numeric)
+S3method(as_lazy_tensor,torch_tensor)
 S3method(as_torch_callback,R6ClassGenerator)
 S3method(as_torch_callback,TorchCallback)
 S3method(as_torch_callback,character)
@@ -22,10 +25,16 @@ S3method(as_torch_loss,nn_loss)
 S3method(as_torch_optimizer,TorchOptimizer)
 S3method(as_torch_optimizer,character)
 S3method(as_torch_optimizer,torch_optimizer_generator)
-S3method(c,imageuri)
+S3method(c,lazy_tensor)
 S3method(col_info,DataBackendLazy)
+S3method(format,lazy_tensor)
+S3method(hash_input,lazy_tensor)
+S3method(materialize,data.frame)
+S3method(materialize,lazy_tensor)
+S3method(materialize,list)
 S3method(print,ModelDescriptor)
 S3method(print,TorchIngressToken)
+S3method(print,lazy_tensor)
 S3method(reset_last_layer,AlexNet)
 S3method(reset_last_layer,resnet)
 S3method(t_clbk,"NULL")
@@ -46,6 +55,7 @@ export(CallbackSetHistory)
 export(CallbackSetProgress)
 export(ContextTorch)
 export(DataBackendLazy)
+export(DataDescriptor)
 export(LearnerTorch)
 export(LearnerTorchAlexNet)
 export(LearnerTorchFeatureless)
@@ -54,6 +64,7 @@ export(LearnerTorchMLP)
 export(LearnerTorchModel)
 export(ModelDescriptor)
 export(PipeOpModule)
+export(PipeOpTaskPreprocTorch)
 export(PipeOpTorch)
 export(PipeOpTorchAvgPool1D)
 export(PipeOpTorchAvgPool2D)
@@ -80,7 +91,7 @@ export(PipeOpTorchHardTanh)
 export(PipeOpTorchHead)
 export(PipeOpTorchIngress)
 export(PipeOpTorchIngressCategorical)
-export(PipeOpTorchIngressImage)
+export(PipeOpTorchIngressLazyTensor)
 export(PipeOpTorchIngressNumeric)
 export(PipeOpTorchLayerNorm)
 export(PipeOpTorchLeakyReLU)
@@ -120,14 +131,19 @@ export(TorchDescriptor)
 export(TorchIngressToken)
 export(TorchLoss)
 export(TorchOptimizer)
+export(as_lazy_tensor)
 export(as_torch_callback)
 export(as_torch_callbacks)
 export(as_torch_loss)
 export(as_torch_optimizer)
+export(assert_lazy_tensor)
+export(auto_device)
 export(batchgetter_categ)
 export(batchgetter_num)
 export(callback_set)
-export(imageuri)
+export(is_lazy_tensor)
+export(lazy_tensor)
+export(materialize)
 export(mlr3torch_callbacks)
 export(mlr3torch_losses)
 export(mlr3torch_optimizers)
@@ -141,6 +157,7 @@ export(nn_merge_sum)
 export(nn_reshape)
 export(nn_squeeze)
 export(nn_unsqueeze)
+export(pipeop_preproc_torch)
 export(reset_last_layer)
 export(t_clbk)
 export(t_clbks)
@@ -150,6 +167,7 @@ export(t_opt)
 export(t_opts)
 export(task_dataset)
 export(torch_callback)
+exportPattern("^PipeOpPreprocTorch")
 import(checkmate)
 import(data.table)
 import(mlr3)

R/CallbackSetHistory.R

@@ -70,9 +70,9 @@ CallbackSetHistory = R6Class("CallbackSetHistory",
         stopf("No eligible measures to plot for set '%s'.", set)
       }
 
-      epoch = score = measure = NULL
+      epoch = score = measure = .data = NULL
       if (ncol(data) == 2L) {
-        ggplot2::ggplot(data = data, ggplot2::aes(x = epoch, y = !!rlang::sym(measures))) +
+        ggplot2::ggplot(data = data, ggplot2::aes(x = epoch, y = .data[[measures]])) +
           ggplot2::geom_line() +
           ggplot2::geom_point() +
           ggplot2::labs(

R/DataBackendLazy.R

@@ -21,15 +21,16 @@
 #'
 #' Beware that accessing the backend's hash also contructs the backend.
 #'
+#' Note that while in most cases the data contains [`lazy_tensor`] columns, this is not necessary and the naming
+#' of this class has nothing to do with the [`lazy_tensor`] data type.
+#'
 #' **Important**
 #'
 #' When the constructor generates `factor()` variables it is important that the ordering of the levels in data
 #' corresponds to the ordering of the levels in the `col_info` argument.
-#' Because the ordering of the level depends on the locale, it is recommended to e.g. use the `C` locale in the
-#' `constructor` function, e.g. with `withr::with_locale()`.
 #'
-#' @param constructor (`function()`)\cr
-#'   A function with no arguments, whose return value must be the actual backend.
+#' @param constructor (`function`)\cr
+#'   A function with argument `backend` (the lazy backend), whose return value must be the actual backend.
 #'   This function is called the first time the field `$backend` is accessed.
 #' @param rownames (`integer()`)\cr
 #'   The row names. Must be a permutation of the rownames of the lazily constructed backend.
@@ -53,7 +54,7 @@
 #' @export
 #' @examples
 #' # We first define a backend constructor
-#' constructor = function() {
+#' constructor = function(backend) {
 #'   cat("Data is constructed!\n")
 #'   DataBackendDataTable$new(
 #'     data.table(x = rnorm(10), y = rnorm(10), row_id = 1:10),
@@ -100,7 +101,7 @@ DataBackendLazy = R6Class("DataBackendLazy",
       assert_choice(primary_key, col_info$id)
       private$.colnames = col_info$id
       assert_choice(primary_key, col_info$id)
-      private$.constructor = assert_function(constructor, nargs = 0)
+      private$.constructor = assert_function(constructor, args = "backend")
 
       super$initialize(data = NULL, primary_key = primary_key, data_formats = data_formats)
     },
@@ -172,7 +173,7 @@ DataBackendLazy = R6Class("DataBackendLazy",
     backend = function(rhs) {
       assert_ro_binding(rhs)
       if (is.null(private$.backend)) {
-        private$.backend = assert_backend(private$.constructor())
+        private$.backend = assert_backend(private$.constructor(self))
 
         f = function(test, x, y, var_name) {
           if (!test(x, y)) {
@@ -184,8 +185,7 @@ DataBackendLazy = R6Class("DataBackendLazy",
           }
         }
 
-        # test_equal does not exist so we abuse test_permutation for primary_key
-        f(test_permutation, private$.backend$primary_key, self$primary_key, "primary key")
+        f(identical, private$.backend$primary_key, self$primary_key, "primary key")
         f(test_permutation, private$.backend$rownames, self$rownames, "row identifiers")
         f(test_permutation, private$.backend$colnames, private$.colnames, "column names")
         f(test_equal_col_info, col_info(private$.backend), private$.col_info, "column information")