[Breaking] Fix custom metric for multi output. (#5954)

* Set output margin to true for custom metric.  This fixes only R and Python.

R-package/R/utils.R

@@ -173,7 +173,7 @@ xgb.iter.eval <- function(booster_handle, watchlist, iter, feval = NULL) {
   } else {
     res <- sapply(seq_along(watchlist), function(j) {
       w <- watchlist[[j]]
-      preds <- predict(booster_handle, w, ntreelimit = 0) # predict using all trees
+      preds <- predict(booster_handle, w, outputmargin = TRUE, ntreelimit = 0) # predict using all trees
       eval_res <- feval(preds, w)
       out <- eval_res$value
       names(out) <- paste0(evnames[j], "-", eval_res$metric)

R-package/tests/testthat/test_custom_objective.R

@@ -79,6 +79,10 @@ test_that("custom objective with multi-class works", {
     hess <- rnorm(dim(as.matrix(preds))[1])
     return (list(grad = grad, hess = hess))
   }
+  fake_merror <- function(preds, dtrain) {
+    expect_equal(dim(data)[1] * nclasses, dim(as.matrix(preds))[1])
+  }
   param$objective <- fake_softprob
+  param$eval_metric <- fake_merror
   bst <- xgb.train(param, dtrain, 1, num_class = nclasses)
 })

demo/guide-python/custom_softmax.py

@@ -75,7 +75,7 @@ def softprob_obj(predt: np.ndarray, data: xgb.DMatrix):
     return grad, hess
 
 
-def predict(booster, X):
+def predict(booster: xgb.Booster, X):
     '''A customized prediction function that converts raw prediction to
     target class.
 
@@ -93,15 +93,34 @@ def predict(booster, X):
     return out
 
 
+def merror(predt: np.ndarray, dtrain: xgb.DMatrix):
+    y = dtrain.get_label()
+    # Like custom objective, the predt is untransformed leaf weight
+    assert predt.shape == (kRows, kClasses)
+    out = np.zeros(kRows)
+    for r in range(predt.shape[0]):
+        i = np.argmax(predt[r])
+        out[r] = i
+
+    assert y.shape == out.shape
+
+    errors = np.zeros(kRows)
+    errors[y != out] = 1.0
+    return 'PyMError', np.sum(errors) / kRows
+
+
 def plot_history(custom_results, native_results):
     fig, axs = plt.subplots(2, 1)
     ax0 = axs[0]
     ax1 = axs[1]
 
+    pymerror = custom_results['train']['PyMError']
+    merror = native_results['train']['merror']
+
     x = np.arange(0, kRounds, 1)
-    ax0.plot(x, custom_results['train']['merror'], label='Custom objective')
+    ax0.plot(x, pymerror, label='Custom objective')
     ax0.legend()
-    ax1.plot(x, native_results['train']['merror'], label='multi:softmax')
+    ax1.plot(x, merror, label='multi:softmax')
     ax1.legend()
 
     plt.show()
@@ -110,10 +129,12 @@ def plot_history(custom_results, native_results):
 def main(args):
     custom_results = {}
     # Use our custom objective function
-    booster_custom = xgb.train({'num_class': kClasses},
+    booster_custom = xgb.train({'num_class': kClasses,
+                                'disable_default_eval_metric': True},
                                m,
                                num_boost_round=kRounds,
                                obj=softprob_obj,
+                               feval=merror,
                                evals_result=custom_results,
                                evals=[(m, 'train')])
 
@@ -131,6 +152,8 @@ def main(args):
     # We are reimplementing the loss function in XGBoost, so it should
     # be the same for normal cases.
     assert np.all(predt_custom == predt_native)
+    np.testing.assert_allclose(custom_results['train']['PyMError'],
+                               native_results['train']['merror'])
 
     if args.plot != 0:
         plot_history(custom_results, native_results)

doc/parameter.rst

@@ -40,9 +40,9 @@ General Parameters
 
   - Number of parallel threads used to run XGBoost
 
-* ``disable_default_eval_metric`` [default=0]
+* ``disable_default_eval_metric`` [default=``false``]
 
-  - Flag to disable default metric. Set to >0 to disable.
+  - Flag to disable default metric. Set to 1 or ``true`` to disable.
 
 * ``num_pbuffer`` [set automatically by XGBoost, no need to be set by user]
 

python-package/xgboost/core.py

@@ -1228,7 +1228,8 @@ def eval_set(self, evals, iteration=0, feval=None):
         res = msg.value.decode()
         if feval is not None:
             for dmat, evname in evals:
-                feval_ret = feval(self.predict(dmat, training=False), dmat)
+                feval_ret = feval(self.predict(dmat, training=False,
+                                               output_margin=True), dmat)
                 if isinstance(feval_ret, list):
                     for name, val in feval_ret:
                         res += '\t%s-%s:%f' % (evname, name, val)

src/learner.cc

@@ -154,9 +154,9 @@ LearnerModelParam::LearnerModelParam(
 
 struct LearnerTrainParam : public XGBoostParameter<LearnerTrainParam> {
   // data split mode, can be row, col, or none.
-  DataSplitMode dsplit;
+  DataSplitMode dsplit {DataSplitMode::kAuto};
   // flag to disable default metric
-  int disable_default_eval_metric;
+  bool disable_default_eval_metric {false};
   // FIXME(trivialfis): The following parameters belong to model itself, but can be
   // specified by users.  Move them to model parameter once we can get rid of binary IO.
   std::string booster;
@@ -171,7 +171,7 @@ struct LearnerTrainParam : public XGBoostParameter<LearnerTrainParam> {
         .add_enum("row", DataSplitMode::kRow)
         .describe("Data split mode for distributed training.");
     DMLC_DECLARE_FIELD(disable_default_eval_metric)
-        .set_default(0)
+        .set_default(false)
         .describe("Flag to disable default metric. Set to >0 to disable");
     DMLC_DECLARE_FIELD(booster)
         .set_default("gbtree")
@@ -253,7 +253,7 @@ class LearnerConfiguration : public Learner {
   void Configure() override {
     // Varient of double checked lock
     if (!this->need_configuration_) { return; }
-    std::lock_guard<std::mutex> gard(config_lock_);
+    std::lock_guard<std::mutex> guard(config_lock_);
     if (!this->need_configuration_) { return; }
 
     monitor_.Start("Configure");

tests/python/test_early_stopping.py

@@ -37,11 +37,11 @@ def test_early_stopping_nonparallel(self):
                  eval_set=[(X_test, y_test)])
         assert clf3.best_score == 1
 
-    @pytest.mark.skipif(**tm.no_sklearn())
     def evalerror(self, preds, dtrain):
         from sklearn.metrics import mean_squared_error
 
         labels = dtrain.get_label()
+        preds = 1.0 / (1.0 + np.exp(-preds))
         return 'rmse', mean_squared_error(labels, preds)
 
     @staticmethod