GH-15697 glm mojo offset fold predict error (#16159)

* GH-15697: add R test to reproduce error.

* add generic model to test as well.
* Add fold column info to generic model.
* add compareFrame to compare prediction results.
* add prediction with fold column removed

Co-authored-by: wendycwong <[email protected]>
Co-authored-by: Veronika Maurerová <[email protected]>

h2o-genmodel/src/main/java/hex/genmodel/AbstractMojoWriter.java

@@ -200,6 +200,9 @@ private void addCommonModelInfo() throws IOException {
  if (model.offsetColumn() != null) {
  writekv("offset_column", model.offsetColumn());
  }
+ if (model.foldColumn() != null) {
+ writekv("fold_column", model.foldColumn());
+ }
  writekv("balance_classes", model.balanceClasses());
  writekv("default_threshold", model.defaultThreshold());
  writekv("prior_class_distrib", Arrays.toString(model.priorClassDist()));

h2o-genmodel/src/main/java/hex/genmodel/GenModel.java

@@ -28,6 +28,7 @@ public abstract class GenModel implements IGenModel, IGeneratedModel, Serializab
 
  /** Name of the column with offsets (used for certain types of models). */
  public String _offsetColumn;
+ public String _foldColumn;
 
  /** Name of the column determine treatment group, currently only for UpliftDRF models */
  public String _treatmentColumn;

h2o-genmodel/src/main/java/hex/genmodel/ModelMojoReader.java

@@ -259,6 +259,7 @@ private void readAll(final boolean readModelMetadata) throws IOException {
  _model._priorClassDistrib = readkv("prior_class_distrib");
  _model._modelClassDistrib = readkv("model_class_distrib");
  _model._offsetColumn = readkv("offset_column");
+ _model._foldColumn = readkv("fold_column");
  _model._mojo_version = ((Number) readkv("mojo_version")).doubleValue();
  checkMaxSupportedMojoVersion();
  readModelData();

h2o-genmodel/src/main/java/hex/genmodel/descriptor/ModelDescriptorBuilder.java

@@ -43,6 +43,7 @@ public static class MojoModelDescriptor implements ModelDescriptor, Serializable
  private final double[] _priorClassDistrib;
  private final double[] _modelClassDistrib;
  private final String _offsetColumn;
+ private final String _foldColumn;
  private final String _weightsColumn;
  private final String _treatmentColumn;
  private final String[][] _domains;
@@ -65,6 +66,7 @@ private MojoModelDescriptor(final MojoModel mojoModel, final String fullAlgorith
  _modelClassDistrib = mojoModel._modelClassDistrib;
  _h2oVersion = mojoModel._h2oVersion;
  _offsetColumn = mojoModel._offsetColumn;
+ _foldColumn = mojoModel._foldColumn;
  _domains = mojoModel._domains;
  _origDomains = mojoModel.getOrigDomainValues();
  _names = mojoModel._names;
@@ -116,7 +118,7 @@ public String weightsColumn() {
 
  @Override
  public String foldColumn() {
- return null;
+ return _foldColumn;
  }
 
  @Override

h2o-r/tests/testdir_algos/glm/runit_gh_15697_mojo_offset_cv.R

@@ -0,0 +1,63 @@
+setwd(normalizePath(dirname(R.utils::commandArgs(asValues=TRUE)$"f")))
+source("../../../scripts/h2o-r-test-setup.R")
+library(data.table)
+
+test_glm_mojo_offset_fold_columns <- function() {
+ mt <- as.h2o(mtcars)
+
+ find_offset <- h2o.glm(
+ x = c("cyl", "disp"),
+ y = "mpg",
+ training_frame = mt,
+ lambda = 0
+ )
+
+ h2o.residual_deviance(find_offset)
+ # [1] 270.7403
+
+ # create offset to look exactly like it's in the model (easy comparison)
+ mt$offset <- mt$cyl * find_offset@model$coefficients[["cyl"]]
+ mt$fold <- h2o.kfold_column(data = mt, nfolds = 3, seed = 123)
+
+ # move "cyl" from being modeled to be an offset
+ # build with a fold column
+ mod_w_offset <- h2o.glm(
+ x = c("disp"),
+ y = "mpg",
+ training_frame = mt,
+ offset = "offset",
+ lambda = 0,
+ fold_column = "fold"
+ )
+
+ h2o.residual_deviance(mod_w_offset)
+ # [1] 270.7403 (match as expected)
+
+ # save out models then immediately reimport
+ mojo_path <- h2o.save_mojo(object = mod_w_offset, path = ".")
+ biny_path <- h2o.saveModel(object = mod_w_offset, path = ".")
+
+ mojo <- h2o.import_mojo(mojo_file_path = mojo_path)
+ genericGLM <- h2o.genericModel(mojo_path)
+ biny <- h2o.loadModel(path = biny_path)
+
+
+ predict <- h2o.predict(object = mod_w_offset, newdata = mt)
+ predict0 <- h2o.predict(object = mojo, newdata = mt)
+ predict1 <- h2o.predict(object = biny, newdata = mt)
+ predict2 <- h2o.predict(object = genericGLM, newdata = mt)
+
+ # check all predicts match
+ compareFrames(predict, predict0)
+ compareFrames(predict, predict1)
+ compareFrames(predict, predict2)
+
+ # remove the fold column in the dataset for prediction
+ predict3 <- h2o.predict(object = mojo, newdata = mt[-c(13)])
+ predict4 <- h2o.predict(object = genericGLM, newdata = mt[-c(13)])
+ # make sure prediction still matches after removing fold column when calling predict
+ compareFrames(predict, predict3)
+ compareFrames(predict, predict4)
+}
+
+doTest("Fix GLM mojo with offset and fold column", test_glm_mojo_offset_fold_columns)