merge bundling

Author: sebffischer

DESCRIPTION

@@ -52,7 +52,7 @@ Imports:
     data.table,
     digest,
     lgr,
-    mlr3 (>= 0.6.0),
+    mlr3 (>= 0.19.0),
     mlr3misc (>= 0.9.0),
     paradox,
     R6,

NAMESPACE

@@ -12,6 +12,10 @@ S3method(as_pipeop,Learner)
 S3method(as_pipeop,PipeOp)
 S3method(as_pipeop,default)
 S3method(disable_internal_tuning,GraphLearner)
+S3method(marshal_model,Multiplicity)
+S3method(marshal_model,graph_learner_model)
+S3method(marshal_model,pipeop_impute_learner_state)
+S3method(marshal_model,pipeop_learner_cv_state)
 S3method(po,"NULL")
 S3method(po,Filter)
 S3method(po,Learner)
@@ -25,6 +29,10 @@ S3method(predict,Graph)
 S3method(print,Multiplicity)
 S3method(print,Selector)
 S3method(set_validate,GraphLearner)
+S3method(unmarshal_model,Multiplicity_marshaled)
+S3method(unmarshal_model,graph_learner_model_marshaled)
+S3method(unmarshal_model,pipeop_impute_learner_state_marshaled)
+S3method(unmarshal_model,pipeop_learner_cv_state_marshaled)
 export("%>>!%")
 export("%>>%")
 export(Graph)
@@ -155,5 +163,6 @@ importFrom(data.table,as.data.table)
 importFrom(digest,digest)
 importFrom(stats,setNames)
 importFrom(utils,bibentry)
+importFrom(utils,head)
 importFrom(utils,tail)
 importFrom(withr,with_options)

NEWS.md

@@ -6,6 +6,8 @@
 * Minor documentation fixes.
 * Test helpers are now available in `inst/`. These are considered experimental and unstable.
 
+* Added marshaling support to `GraphLearner`
+
 # mlr3pipelines 0.5.1
 
 * Changed the ID of `PipeOpFeatureUnion` used in `ppl("robustify")` and `ppl("stacking")`.

R/GraphLearner.R

@@ -56,7 +56,16 @@
 #' * `validate` :: `numeric(1)`, `"predefined"`, `"test"` or `NULL`\cr
 #'   How to construct the validation data. This also has to be configured in the individual learners wrapped by
 #'   `PipeOpLearner`, see [`set_validate.GraphLearner`] on how to configure this.
+#' * `marshaled` :: `logical(1)`\cr
+#'   Whether the learner is marshaled.
 #'
+#' @section Methods:
+#' * `marshal(...)`\cr
+#'   (any) -> `self`\cr
+#'   Marshal the model.
+#' * `unmarshal(...)`\cr
+#'   (any) -> `self`\cr
+#'   Unmarshal the model.
 #'
 #' @section Internals:
 #' [`as_graph()`] is called on the `graph` argument, so it can technically also be a `list` of things, which is
@@ -150,6 +159,12 @@ GraphLearner = R6Class("GraphLearner", inherit = Learner,
         if (length(last_pipeop_id) == 0) stop("No Learner PipeOp found.")
       }
       learner_model$base_learner(recursive - 1)
+    },
+    marshal = function(...) {
+      learner_marshal(.learner = self, ...)
+    },
+    unmarshal = function(...) {
+      learner_unmarshal(.learner = self, ...)
     }
   ),
   active = list(
@@ -169,7 +184,9 @@ GraphLearner = R6Class("GraphLearner", inherit = Learner,
         private$.validate = assert_validate(rhs)
       }
       private$.validate
-
+    },
+    marshaled = function() {
+      learner_marshaled(self)
     },
     hash = function() {
       digest(list(class(self), self$id, self$graph$hash, private$.predict_type, private$.validate,
@@ -260,6 +277,7 @@ GraphLearner = R6Class("GraphLearner", inherit = Learner,
       on.exit({self$graph$state = NULL})
       self$graph$train(task)
       state = self$graph$state
+      class(state) = c("graph_learner_model", class(state))
       state
     },
     .predict = function(task) {
@@ -414,6 +432,27 @@ disable_internal_tuning.GraphLearner = function(learner, ids, ...) {
 }
 
 
+#' @export
+marshal_model.graph_learner_model = function(model, inplace = FALSE, ...) {
+  xm = map(.x = model, .f = marshal_model, inplace = inplace, ...)
+  # if none of the states required any marshaling we return the model as-is
+  if (!some(xm, is_marshaled_model)) return(model)
+
+  structure(list(
+    marshaled = xm,
+    packages = "mlr3pipelines"
+  ), class = c("graph_learner_model_marshaled", "list_marshaled", "marshaled"))
+}
+
+#' @export
+unmarshal_model.graph_learner_model_marshaled = function(model, inplace = FALSE, ...) {
+  # need to re-create the class as it gets lost during marshaling
+  structure(
+    map(.x = model$marshaled, .f = unmarshal_model, inplace = inplace, ...),
+    class = gsub(x = head(class(model), n = -1), pattern = "_marshaled$", replacement = "")
+  )
+}
+
 #' @export
 as_learner.Graph = function(x, clone = FALSE, ...) {
   GraphLearner$new(x, clone_graph = clone)

R/PipeOpImputeLearner.R

@@ -44,6 +44,8 @@
 #' for each column. If a column consists of missing values only during training, the `model` is `0` or the levels of the
 #' feature; these are used for sampling during prediction.
 #'
+#' This state is given the class `"pipeop_impute_learner_state"`.
+#'
 #' @section Parameters:
 #' The parameters are the parameters inherited from [`PipeOpImpute`], in addition to the parameters of the [`Learner`][mlr3::Learner]
 #' used for imputation.
@@ -116,6 +118,13 @@ PipeOpImputeLearner = R6Class("PipeOpImputeLearner",
       )
       super$initialize(id, param_set = alist(private$.learner$param_set), param_vals = param_vals,
         whole_task_dependent = TRUE, feature_types = feature_types)
+    },
+    train = function(inputs) {
+      outputs = super$train(inputs)
+      self$state = multiplicity_recurse(self$state, function(state) {
+        structure(state, class = c("pipeop_impute_learner_state", class(state)))
+      })
+      return(outputs)
     }
   ),
   active = list(
@@ -206,3 +215,25 @@ mlr_pipeops$add("imputelearner", PipeOpImputeLearner, list(R6Class("Learner", pu
 convert_to_task = function(id = "imputing", data, target, task_type, ...) {
   get(mlr_reflections$task_types[task_type, mult = "first"]$task)$new(id = id, backend = data, target = target, ...)
 }
+
+#' @export
+marshal_model.pipeop_impute_learner_state = function(model, inplace = FALSE, ...) {
+  prev_class = class(model)
+  model$model = map(model$model, marshal_model, inplace = inplace, ...)
+
+  if (!some(model$model, is_marshaled_model)) {
+    return(model)
+  }
+
+  structure(
+    list(marshaled = model, packages = "mlr3pipelines"),
+    class = c(paste0(prev_class, "_marshaled"), "marshaled")
+  )
+}
+
+#' @export
+unmarshal_model.pipeop_impute_learner_state_marshaled = function(model, inplace = FALSE, ...) {
+  state = model$marshaled
+  state$model = map(state$model, unmarshal_model, inplace = inplace, ...)
+  return(state)
+}

R/PipeOpLearner.R

@@ -94,8 +94,7 @@ PipeOpLearner = R6Class("PipeOpLearner", inherit = PipeOp,
       super$initialize(id, param_set = alist(private$.learner$param_set), param_vals = param_vals,
         input = data.table(name = "input", train = task_type, predict = task_type),
         output = data.table(name = "output", train = "NULL", predict = out_type),
-        tags = "learner", packages = learner$packages
-      )
+        tags = "learner", packages = learner$packages)
     }
   ),
   active = list(

R/PipeOpLearnerCV.R

@@ -61,6 +61,8 @@
 #' * `predict_time` :: `NULL` | `numeric(1)`
 #'   Prediction time, in seconds.
 #'
+#' This state is given the class `"pipeop_learner_cv_state"`.
+#'
 #' @section Parameters:
 #' The parameters are the parameters inherited from the [`PipeOpTaskPreproc`], as well as the parameters of the [`Learner`][mlr3::Learner] wrapped by this object.
 #' Besides that, parameters introduced are:
@@ -144,8 +146,14 @@ PipeOpLearnerCV = R6Class("PipeOpLearnerCV",
       # private$.crossval_param_set$add_dep("folds", "method", CondEqual$new("cv"))  # don't do this.
 
       super$initialize(id, alist(resampling = private$.crossval_param_set, private$.learner$param_set), param_vals = param_vals, can_subset_cols = TRUE, task_type = task_type, tags = c("learner", "ensemble"))
+    },
+    train = function(inputs) {
+      outputs = super$train(inputs)
+      self$state = multiplicity_recurse(self$state, function(state) {
+          structure(state, class = c("pipeop_learner_cv_state", class(state)))
+      })
+      return(outputs)
     }
-
   ),
   active = list(
     learner = function(val) {
@@ -224,4 +232,29 @@ PipeOpLearnerCV = R6Class("PipeOpLearnerCV",
   )
 )
 
+#' @export
+marshal_model.pipeop_learner_cv_state = function(model, inplace = FALSE, ...) {
+  # Note that a Learner state contains other reference objects, but we don't clone them here, even when inplace
+  # is FALSE. For our use-case this is just not necessary and would cause unnecessary overhead in the mlr3
+  # workhorse function
+  model$model = marshal_model(model$model, inplace = inplace)
+  # only wrap this in a marshaled class if the model was actually marshaled above
+  # (the default marshal method does nothing)
+  if (is_marshaled_model(model$model)) {
+    model = structure(
+      list(marshaled = model, packages = "mlr3pipelines"),
+      class = c(paste0(class(model), "_marshaled"), "marshaled")
+    )
+  }
+  model
+}
+
+#' @export
+unmarshal_model.pipeop_learner_cv_state_marshaled = function(model, inplace = FALSE, ...) {
+  state_marshaled = model$marshaled
+  state_marshaled$model = unmarshal_model(state_marshaled$model, inplace = inplace)
+  state_marshaled
+}
+
+
 mlr_pipeops$add("learner_cv", PipeOpLearnerCV, list(R6Class("Learner", public = list(id = "learner_cv", task_type = "classif", param_set = ps()))$new()))

R/PipeOpTaskPreproc.R

@@ -187,8 +187,7 @@ PipeOpTaskPreproc = R6Class("PipeOpTaskPreproc",
       super$initialize(id = id, param_set = param_set, param_vals = param_vals,
         input = data.table(name = "input", train = task_type, predict = task_type),
         output = data.table(name = "output", train = task_type, predict = task_type),
-        packages = packages, tags = c(tags, "data transform")
-      )
+        packages = packages, tags = c(tags, "data transform"))
     }
   ),
   active = list(

R/multiplicity.R

@@ -115,3 +115,16 @@ multiplicity_nests_deeper_than = function(x, cutoff) {
   }
   ret
 }
+
+#' @export
+marshal_model.Multiplicity = function(model, inplace = FALSE, ...) {
+  structure(list(
+    marshaled = multiplicity_recurse(model, marshal_model, inplace = inplace, ...),
+    packages = "mlr3pipelines"
+  ), class = c("Multiplicity_marshaled", "marshaled"))
+}
+
+#' @export
+unmarshal_model.Multiplicity_marshaled = function(model, inplace = FALSE, ...) {
+  multiplicity_recurse(model$marshaled, unmarshal_model, inplace = inplace, ...)
+}

R/zzz.R

@@ -4,7 +4,7 @@
 #' @import paradox
 #' @import mlr3misc
 #' @importFrom R6 R6Class
-#' @importFrom utils tail
+#' @importFrom utils tail head
 #' @importFrom digest digest
 #' @importFrom withr with_options
 #' @importFrom stats setNames