fix: Correct hash implementations for PipeOps (#372)

sebffischer · web-flow · commit f861ee545048 · 2025-03-27T11:49:15.000+01:00
diff --git a/R/PipeOpTorchAdaptiveAvgPool.R b/R/PipeOpTorchAdaptiveAvgPool.R
@@ -18,6 +18,9 @@ PipeOpTorchAdaptiveAvgPool = R6Class("PipeOpTorchAdaptiveAvgPool",
     }
   ),
   private = list(
+    .additional_phash_input = function() {
+      list(private$.d)
+    },
     .shapes_out = function(shapes_in, param_vals, task) {
       list(adaptive_avg_output_shape(
         shape_in = shapes_in[[1]],
diff --git a/R/PipeOpTorchAvgPool.R b/R/PipeOpTorchAvgPool.R
@@ -31,6 +31,9 @@ PipeOpTorchAvgPool = R6Class("PipeOpTorchAvgPool",
     }
   ),
   private = list(
+    .additional_phash_input = function() {
+      list(private$.d)
+    },
     .shapes_out = function(shapes_in, param_vals, task) {
       list(avg_output_shape(
         shape_in = shapes_in[[1]],
diff --git a/R/PipeOpTorchBatchNorm.R b/R/PipeOpTorchBatchNorm.R
@@ -30,6 +30,9 @@ PipeOpTorchBatchNorm = R6Class("PipeOpTorchBatchNorm",
   private = list(
     .min_dim = NULL,
     .max_dim = NULL,
+    .additional_phash_input = function() {
+      list(private$.min_dim, private$.max_dim)
+    },
     .shapes_out = function(shapes_in, param_vals, task) {
       list(assert_numeric(shapes_in[[1]], min.len = private$.min_dim, max.len = private$.max_dim))
     },
@@ -63,7 +66,6 @@ PipeOpTorchBatchNorm = R6Class("PipeOpTorchBatchNorm",
 #' @template pipeop_torch
 #' @template pipeop_torch_example
 #'
-#'
 #' @export
 PipeOpTorchBatchNorm1D = R6Class("PipeOpTorchBatchNorm1D", inherit = PipeOpTorchBatchNorm,
   public = list(
diff --git a/R/PipeOpTorchConv.R b/R/PipeOpTorchConv.R
@@ -26,6 +26,9 @@ PipeOpTorchConv = R6Class("PipeOpTorchConv",
     }
   ),
   private = list(
+    .additional_phash_input = function() {
+      list(private$.d)
+    },
     .shapes_out = function(shapes_in, param_vals, task) {
       list(conv_output_shape(
         shape_in = shapes_in[[1]],
diff --git a/R/PipeOpTorchConvTranspose.R b/R/PipeOpTorchConvTranspose.R
@@ -35,6 +35,9 @@ PipeOpTorchConvTranspose = R6Class("PipeOpTorchConvTranspose",
     }
   ),
   private = list(
+    .additional_phash_input = function() {
+      list(private$.d)
+    },
     .shapes_out = function(shapes_in, param_vals, task) {
       list(conv_transpose_output_shape(
         shape_in = shapes_in[[1]],
diff --git a/R/PipeOpTorchMaxPool.R b/R/PipeOpTorchMaxPool.R
@@ -31,6 +31,9 @@ PipeOpTorchMaxPool = R6Class("PipeOpTorchMaxPool",
     }
   ),
   private = list(
+    .additional_phash_input = function() {
+      list(d = private$.d)
+    },
     .shapes_out = function(shapes_in, param_vals, task) {
       res = list(max_output_shape(
         shape_in = shapes_in[[1]],
diff --git a/R/PipeOpTorchMerge.R b/R/PipeOpTorchMerge.R
@@ -45,6 +45,9 @@ PipeOpTorchMerge = R6Class("PipeOpTorchMerge",
   ),
   private = list(
     .innum = NULL,
+    .additional_phash_input = function() {
+      list(private$.innum)
+    },
     .shapes_out = function(shapes_in, param_vals, task) {
       # note that this slightly deviates from the actual broadcasting rules implemented by torch, i.e. we don't fill
       # up missing dimension with 1s because the first dimension is usually the batch dimension.
diff --git a/tests/testthat/helper_autotest.R b/tests/testthat/helper_autotest.R
@@ -28,6 +28,10 @@ expect_pipeop_torch = function(graph, id, task, module_class = id, exclude_args
   result = graph$train(task)
   md = result[[1]]
 
+  # PipeOp overwrites hash and phash if necessary
+  testthat::expect_warning(po_test$hash, regexp = NA)
+  testthat::expect_warning(po_test$phash, regexp = NA)
+
   modulegraph = md$graph
   po_module = modulegraph$pipeops[[id]]
   if (is.null(po_module$module)) {
@@ -325,6 +329,10 @@ expect_pipeop_torch_preprocess = function(obj, shapes_in, exclude = character(0)
   if (is.null(seed)) {
     seed = sample.int(100000, 1)
   }
+  # overwrites .additional_phash_input where necessary
+  testthat::expect_warning(obj$hash, regexp = NA)
+  testthat::expect_warning(obj$phash, regexp = NA)
+
   expect_pipeop(obj)
   expect_class(obj, "PipeOpTaskPreprocTorch")
   # a) Check that all parameters but stages have tags train and predict (this should hold in basically all cases)
@@ -423,6 +431,11 @@ expect_pipeop_torch_preprocess = function(obj, shapes_in, exclude = character(0)
 
 expect_learner_torch = function(learner, task, check_man = TRUE, check_id = TRUE) {
   checkmate::expect_class(learner, "LearnerTorch")
+
+  # overwrites .additional_phash_input where necessary
+  testthat::expect_warning(learner$hash, regexp = NA)
+  testthat::expect_warning(learner$phash, regexp = NA)
+
   get("expect_learner", envir = .GlobalEnv)(learner)
   # state cloning is tested separately
   learner1 = learner
