[pyspark] Use quantile dmatrix. (#8284)

doc/parameter.rst

@@ -349,7 +349,7 @@ Specify the learning task and the corresponding learning objective. The objectiv
   - ``reg:squaredlogerror``: regression with squared log loss :math:`\frac{1}{2}[log(pred + 1) - log(label + 1)]^2`.  All input labels are required to be greater than -1.  Also, see metric ``rmsle`` for possible issue  with this objective.
   - ``reg:logistic``: logistic regression.
   - ``reg:pseudohubererror``: regression with Pseudo Huber loss, a twice differentiable alternative to absolute loss.
-  - ``reg:absoluteerror``: Regression with L1 error. When tree model is used, leaf value is refreshed after tree construction.
+  - ``reg:absoluteerror``: Regression with L1 error. When tree model is used, leaf value is refreshed after tree construction. If used in distributed training, the leaf value is calculated as the mean value from all workers, which is not guaranteed to be optimal.
   - ``binary:logistic``: logistic regression for binary classification, output probability
   - ``binary:logitraw``: logistic regression for binary classification, output score before logistic transformation
   - ``binary:hinge``: hinge loss for binary classification. This makes predictions of 0 or 1, rather than producing probabilities.

python-package/xgboost/core.py

@@ -105,6 +105,11 @@ def from_cstr_to_pystr(data: CStrPptr, length: c_bst_ulong) -> List[str]:
     return res
 
 
+def make_jcargs(**kwargs: Any) -> bytes:
+    "Make JSON-based arguments for C functions."
+    return from_pystr_to_cstr(json.dumps(kwargs))
+
+
 IterRange = TypeVar("IterRange", Optional[Tuple[int, int]], Tuple[int, int])
 
 
@@ -1256,7 +1261,7 @@ class _ProxyDMatrix(DMatrix):
     def _set_data_from_cuda_interface(self, data: DataType) -> None:
         """Set data from CUDA array interface."""
         interface = data.__cuda_array_interface__
-        interface_str = bytes(json.dumps(interface, indent=2), "utf-8")
+        interface_str = bytes(json.dumps(interface), "utf-8")
         _check_call(
             _LIB.XGProxyDMatrixSetDataCudaArrayInterface(self.handle, interface_str)
         )
@@ -1357,6 +1362,26 @@ class QuantileDMatrix(DMatrix):
                 "Only one of the eval_qid or eval_group for each evaluation "
                 "dataset should be provided."
             )
+        if isinstance(data, DataIter):
+            if any(
+                info is not None
+                for info in (
+                    label,
+                    weight,
+                    base_margin,
+                    feature_names,
+                    feature_types,
+                    group,
+                    qid,
+                    label_lower_bound,
+                    label_upper_bound,
+                    feature_weights,
+                )
+            ):
+                raise ValueError(
+                    "If data iterator is used as input, data like label should be "
+                    "specified as batch argument."
+                )
 
         self._init(
             data,
@@ -1405,12 +1430,9 @@ class QuantileDMatrix(DMatrix):
                 "in iterator to fix this error."
             )
 
-        args = {
+        config = make_jcargs(
-            "nthread": self.nthread,
+            nthread=self.nthread, missing=self.missing, max_bin=self.max_bin
-            "missing": self.missing,
+        )
-            "max_bin": self.max_bin,
-        }
-        config = from_pystr_to_cstr(json.dumps(args))
         ret = _LIB.XGQuantileDMatrixCreateFromCallback(
             None,
             it.proxy.handle,
@@ -2375,7 +2397,7 @@ class Booster:
         """
         length = c_bst_ulong()
         cptr = ctypes.POINTER(ctypes.c_char)()
-        config = from_pystr_to_cstr(json.dumps({"format": raw_format}))
+        config = make_jcargs(format=raw_format)
         _check_call(
             _LIB.XGBoosterSaveModelToBuffer(
                 self.handle, config, ctypes.byref(length), ctypes.byref(cptr)
@@ -2570,9 +2592,6 @@ class Booster:
         `n_classes`, otherwise they're scalars.
         """
         fmap = os.fspath(os.path.expanduser(fmap))
-        args = from_pystr_to_cstr(
-            json.dumps({"importance_type": importance_type, "feature_map": fmap})
-        )
         features = ctypes.POINTER(ctypes.c_char_p)()
         scores = ctypes.POINTER(ctypes.c_float)()
         n_out_features = c_bst_ulong()
@@ -2582,7 +2601,7 @@ class Booster:
         _check_call(
             _LIB.XGBoosterFeatureScore(
                 self.handle,
-                args,
+                make_jcargs(importance_type=importance_type, feature_map=fmap),
                 ctypes.byref(n_out_features),
                 ctypes.byref(features),
                 ctypes.byref(out_dim),

python-package/xgboost/dask.py

@@ -573,6 +573,7 @@ class DaskPartitionIter(DataIter):  # pylint: disable=R0902
         label_upper_bound: Optional[List[Any]] = None,
         feature_names: Optional[FeatureNames] = None,
         feature_types: Optional[Union[Any, List[Any]]] = None,
+        feature_weights: Optional[Any] = None,
     ) -> None:
         self._data = data
         self._label = label
@@ -583,6 +584,7 @@ class DaskPartitionIter(DataIter):  # pylint: disable=R0902
         self._label_upper_bound = label_upper_bound
         self._feature_names = feature_names
         self._feature_types = feature_types
+        self._feature_weights = feature_weights
 
         assert isinstance(self._data, collections.abc.Sequence)
 
@@ -633,6 +635,7 @@ class DaskPartitionIter(DataIter):  # pylint: disable=R0902
             label_upper_bound=self._get("_label_upper_bound"),
             feature_names=feature_names,
             feature_types=self._feature_types,
+            feature_weights=self._feature_weights,
         )
         self._iter += 1
         return 1
@@ -731,19 +734,21 @@ def _create_quantile_dmatrix(
         return d
 
     unzipped_dict = _get_worker_parts(parts)
-    it = DaskPartitionIter(**unzipped_dict)
+    it = DaskPartitionIter(
+        **unzipped_dict,
+        feature_types=feature_types,
+        feature_names=feature_names,
+        feature_weights=feature_weights,
+    )
 
     dmatrix = QuantileDMatrix(
         it,
         missing=missing,
-        feature_names=feature_names,
-        feature_types=feature_types,
         nthread=nthread,
         max_bin=max_bin,
         ref=ref,
         enable_categorical=enable_categorical,
     )
-    dmatrix.set_info(feature_weights=feature_weights)
     return dmatrix
 
 

python-package/xgboost/spark/core.py

@@ -747,6 +747,7 @@ class _SparkXGBEstimator(Estimator, _SparkXGBParams, MLReadable, MLWritable):
             k: v for k, v in train_call_kwargs_params.items() if v is not None
         }
         dmatrix_kwargs = {k: v for k, v in dmatrix_kwargs.items() if v is not None}
+        use_qdm = booster_params.get("tree_method") in ("hist", "gpu_hist")
 
         def _train_booster(pandas_df_iter):
             """Takes in an RDD partition and outputs a booster for that partition after
@@ -759,20 +760,17 @@ class _SparkXGBEstimator(Estimator, _SparkXGBParams, MLReadable, MLWritable):
             context.barrier()
 
             gpu_id = None
+
+            if use_qdm and (booster_params.get("max_bin", None) is not None):
+                dmatrix_kwargs["max_bin"] = booster_params["max_bin"]
+
             if use_gpu:
                 gpu_id = context.partitionId() if is_local else _get_gpu_id(context)
                 booster_params["gpu_id"] = gpu_id
 
-                # max_bin is needed for qdm
-                if (
-                    features_cols_names is not None
-                    and booster_params.get("max_bin", None) is not None
-                ):
-                    dmatrix_kwargs["max_bin"] = booster_params["max_bin"]
-
             _rabit_args = {}
             if context.partitionId() == 0:
-                get_logger("XGBoostPySpark").info(
+                get_logger("XGBoostPySpark").debug(
                     "booster params: %s\n"
                     "train_call_kwargs_params: %s\n"
                     "dmatrix_kwargs: %s",
@@ -791,6 +789,7 @@ class _SparkXGBEstimator(Estimator, _SparkXGBParams, MLReadable, MLWritable):
                     pandas_df_iter,
                     features_cols_names,
                     gpu_id,
+                    use_qdm,
                     dmatrix_kwargs,
                     enable_sparse_data_optim=enable_sparse_data_optim,
                     has_validation_col=has_validation_col,

python-package/xgboost/spark/data.py

@@ -1,13 +1,13 @@
 """Utilities for processing spark partitions."""
 from collections import defaultdict, namedtuple
-from typing import Any, Callable, Dict, Iterator, List, Optional, Sequence, Tuple
+from typing import Any, Callable, Dict, Iterator, List, Optional, Sequence, Tuple, Union
 
 import numpy as np
 import pandas as pd
 from scipy.sparse import csr_matrix
 from xgboost.compat import concat
 
-from xgboost import DataIter, DeviceQuantileDMatrix, DMatrix
+from xgboost import DataIter, DMatrix, QuantileDMatrix
 
 from .utils import get_logger  # type: ignore
 
@@ -67,10 +67,13 @@ def cache_partitions(
 class PartIter(DataIter):
     """Iterator for creating Quantile DMatrix from partitions."""
 
-    def __init__(self, data: Dict[str, List], device_id: Optional[int]) -> None:
+    def __init__(
+        self, data: Dict[str, List], device_id: Optional[int], **kwargs: Any
+    ) -> None:
         self._iter = 0
         self._device_id = device_id
         self._data = data
+        self._kwargs = kwargs
 
         super().__init__()
 
@@ -98,6 +101,7 @@ class PartIter(DataIter):
             weight=self._fetch(self._data.get(alias.weight, None)),
             base_margin=self._fetch(self._data.get(alias.margin, None)),
             qid=self._fetch(self._data.get(alias.qid, None)),
+            **self._kwargs,
         )
         self._iter += 1
         return 1
@@ -149,24 +153,52 @@ def _read_csr_matrix_from_unwrapped_spark_vec(part: pd.DataFrame) -> csr_matrix:
     )
 
 
+def make_qdm(
+    data: Dict[str, List[np.ndarray]],
+    gpu_id: Optional[int],
+    meta: Dict[str, Any],
+    ref: Optional[DMatrix],
+    params: Dict[str, Any],
+) -> DMatrix:
+    """Handle empty partition for QuantileDMatrix."""
+    if not data:
+        return QuantileDMatrix(np.empty((0, 0)), ref=ref)
+    it = PartIter(data, gpu_id, **meta)
+    m = QuantileDMatrix(it, **params, ref=ref)
+    return m
+
+
 def create_dmatrix_from_partitions(  # pylint: disable=too-many-arguments
     iterator: Iterator[pd.DataFrame],
     feature_cols: Optional[Sequence[str]],
     gpu_id: Optional[int],
+    use_qdm: bool,
     kwargs: Dict[str, Any],  # use dict to make sure this parameter is passed.
     enable_sparse_data_optim: bool,
     has_validation_col: bool,
 ) -> Tuple[DMatrix, Optional[DMatrix]]:
-    """Create DMatrix from spark data partitions. This is not particularly efficient as
+    """Create DMatrix from spark data partitions.
-    we need to convert the pandas series format to numpy then concatenate all the data.
 
     Parameters
     ----------
     iterator :
         Pyspark partition iterator.
+    feature_cols:
+        A sequence of feature names, used only when rapids plugin is enabled.
+    gpu_id:
+        Device ordinal, used when GPU is enabled.
+    use_qdm :
+        Whether QuantileDMatrix should be used instead of DMatrix.
     kwargs :
         Metainfo for DMatrix.
+    enable_sparse_data_optim :
+        Whether sparse data should be unwrapped
+    has_validation:
+        Whether there's validation data.
 
+    Returns
+    -------
+    Training DMatrix and an optional validation DMatrix.
     """
     # pylint: disable=too-many-locals, too-many-statements
     train_data: Dict[str, List[np.ndarray]] = defaultdict(list)
@@ -206,16 +238,16 @@ def create_dmatrix_from_partitions(  # pylint: disable=too-many-arguments
             else:
                 train_data[name].append(array)
 
-    def append_dqm(part: pd.DataFrame, name: str, is_valid: bool) -> None:
+    def append_qdm(part: pd.DataFrame, name: str, is_valid: bool) -> None:
-        """Preprocessing for DeviceQuantileDMatrix"""
+        """Preprocessing for QuantileDMatrix."""
         nonlocal n_features
         if name == alias.data or name in part.columns:
-            if name == alias.data:
+            if name == alias.data and feature_cols is not None:
-                cname = feature_cols
+                array = part[feature_cols]
             else:
-                cname = name
+                array = part[name]
+                array = stack_series(array)
 
-            array = part[cname]
             if name == alias.data:
                 if n_features == 0:
                     n_features = array.shape[1]
@@ -228,6 +260,10 @@ def create_dmatrix_from_partitions(  # pylint: disable=too-many-arguments
 
     def make(values: Dict[str, List[np.ndarray]], kwargs: Dict[str, Any]) -> DMatrix:
         if len(values) == 0:
+            get_logger("XGBoostPySpark").warning(
+                "Detected an empty partition in the training data. Consider to enable"
+                " repartition_random_shuffle"
+            )
             # We must construct an empty DMatrix to bypass the AllReduce
             return DMatrix(data=np.empty((0, 0)), **kwargs)
 
@@ -240,32 +276,62 @@ def create_dmatrix_from_partitions(  # pylint: disable=too-many-arguments
             data=data, label=label, weight=weight, base_margin=margin, qid=qid, **kwargs
         )
 
-    is_dmatrix = feature_cols is None
-    if is_dmatrix:
     if enable_sparse_data_optim:
         append_fn = append_m_sparse
         assert "missing" in kwargs and kwargs["missing"] == 0.0
     else:
         append_fn = append_m
-        cache_partitions(iterator, append_fn)
+
-        if len(train_data) == 0:
+    def split_params() -> Tuple[Dict[str, Any], Dict[str, Union[int, float, bool]]]:
-            get_logger("XGBoostPySpark").warning(
+        # FIXME(jiamingy): we really need a better way to bridge distributed frameworks
-                "Detected an empty partition in the training data. "
+        # to XGBoost native interface and prevent scattering parameters like this.
-                "Consider to enable repartition_random_shuffle"
+
+        # parameters that are not related to data.
+        non_data_keys = (
+            "max_bin",
+            "missing",
+            "silent",
+            "nthread",
+            "enable_categorical",
         )
-        dtrain = make(train_data, kwargs)
+        non_data_params = {}
+        meta = {}
+        for k, v in kwargs.items():
+            if k in non_data_keys:
+                non_data_params[k] = v
             else:
-        cache_partitions(iterator, append_dqm)
+                meta[k] = v
-        it = PartIter(train_data, gpu_id)
+        return meta, non_data_params
-        dtrain = DeviceQuantileDMatrix(it, **kwargs)
+
+    meta, params = split_params()
+
+    if feature_cols is not None:  # rapidsai plugin
+        assert gpu_id is not None
+        assert use_qdm is True
+        cache_partitions(iterator, append_qdm)
+        dtrain: DMatrix = make_qdm(train_data, gpu_id, meta, None, params)
+    elif use_qdm:
+        cache_partitions(iterator, append_qdm)
+        dtrain = make_qdm(train_data, gpu_id, meta, None, params)
+    else:
+        cache_partitions(iterator, append_fn)
+        dtrain = make(train_data, kwargs)
 
     # Using has_validation_col here to indicate if there is validation col
     # instead of getting it from iterator, since the iterator may be empty
     # in some special case. That is to say, we must ensure every worker
-    # construct DMatrix even there is no any data since we need to ensure every
+    # construct DMatrix even there is no data since we need to ensure every
     # worker do the AllReduce when constructing DMatrix, or else it may hang
     # forever.
+    if has_validation_col:
+        if use_qdm:
+            dvalid: Optional[DMatrix] = make_qdm(
+                valid_data, gpu_id, meta, dtrain, params
+            )
+        else:
             dvalid = make(valid_data, kwargs) if has_validation_col else None
+    else:
+        dvalid = None
 
     if dvalid is not None:
         assert dvalid.num_col() == dtrain.num_col()

tests/python-gpu/test_gpu_spark/test_data.py

@@ -20,4 +20,6 @@ from test_spark.test_data import run_dmatrix_ctor
 
 @pytest.mark.skipif(**tm.no_cudf())
 def test_qdm_ctor() -> None:
-    run_dmatrix_ctor(True)
+    run_dmatrix_ctor(is_dqm=True, on_gpu=True)
+    with pytest.raises(AssertionError):
+        run_dmatrix_ctor(is_dqm=False, on_gpu=True)

tests/python-gpu/test_gpu_with_dask/test_gpu_with_dask.py

@@ -188,12 +188,9 @@ def run_gpu_hist(
 
     # See note on `ObjFunction::UpdateTreeLeaf`.
     update_leaf = dataset.name.endswith("-l1")
-    if update_leaf and len(history) == 2:
+    if update_leaf:
         assert history[0] + 1e-2 >= history[-1]
         return
-    if update_leaf and len(history) > 2:
-        assert history[0] >= history[-1]
-        return
     else:
         assert tm.non_increasing(history)
 

tests/python/test_spark/test_data.py

@@ -37,7 +37,7 @@ def test_stack() -> None:
     assert b.shape == (2, 1)
 
 
-def run_dmatrix_ctor(is_dqm: bool) -> None:
+def run_dmatrix_ctor(is_dqm: bool, on_gpu: bool) -> None:
     rng = np.random.default_rng(0)
     dfs: List[pd.DataFrame] = []
     n_features = 16
@@ -57,7 +57,7 @@ def run_dmatrix_ctor(is_dqm: bool) -> None:
         df = pd.DataFrame(
             {alias.label: y, alias.margin: m, alias.weight: w, alias.valid: valid}
         )
-        if is_dqm:
+        if on_gpu:
             for j in range(X.shape[1]):
                 df[f"feat-{j}"] = pd.Series(X[:, j])
         else:
@@ -65,14 +65,18 @@ def run_dmatrix_ctor(is_dqm: bool) -> None:
         dfs.append(df)
 
     kwargs = {"feature_types": feature_types}
-    if is_dqm:
+    if on_gpu:
         cols = [f"feat-{i}" for i in range(n_features)]
         train_Xy, valid_Xy = create_dmatrix_from_partitions(
-            iter(dfs), cols, 0, kwargs, False, True
+            iter(dfs), cols, 0, is_dqm, kwargs, False, True
+        )
+    elif is_dqm:
+        train_Xy, valid_Xy = create_dmatrix_from_partitions(
+            iter(dfs), None, None, True, kwargs, False, True
         )
     else:
         train_Xy, valid_Xy = create_dmatrix_from_partitions(
-            iter(dfs), None, None, kwargs, False, True
+            iter(dfs), None, None, False, kwargs, False, True
         )
 
     assert valid_Xy is not None
@@ -106,7 +110,8 @@ def run_dmatrix_ctor(is_dqm: bool) -> None:
 
 
 def test_dmatrix_ctor() -> None:
-    run_dmatrix_ctor(False)
+    run_dmatrix_ctor(is_dqm=False, on_gpu=False)
+    run_dmatrix_ctor(is_dqm=True, on_gpu=False)
 
 
 def test_read_csr_matrix_from_unwrapped_spark_vec() -> None:

tests/python/test_spark/test_spark_local.py

@@ -1047,7 +1047,11 @@ class XgboostLocalTest(SparkTestCase):
         for row in pred_result:
             assert np.isclose(row.prediction, row.expected_prediction, rtol=1e-3)
 
-    def test_empty_validation_data(self):
+    def test_empty_validation_data(self) -> None:
+        for tree_method in [
+            "hist",
+            "approx",
+        ]:  # pytest.mark conflict with python unittest
             df_train = self.session.createDataFrame(
                 [
                     (Vectors.dense(10.1, 11.2, 11.3), 0, False),
@@ -1059,6 +1063,7 @@ class XgboostLocalTest(SparkTestCase):
             )
             classifier = SparkXGBClassifier(
                 num_workers=2,
+                tree_method=tree_method,
                 min_child_weight=0.0,
                 reg_alpha=0,
                 reg_lambda=0,
@@ -1069,7 +1074,11 @@ class XgboostLocalTest(SparkTestCase):
             for row in pred_result:
                 self.assertEqual(row.prediction, row.label)
 
-    def test_empty_train_data(self):
+    def test_empty_train_data(self) -> None:
+        for tree_method in [
+            "hist",
+            "approx",
+        ]:  # pytest.mark conflict with python unittest
             df_train = self.session.createDataFrame(
                 [
                     (Vectors.dense(10.1, 11.2, 11.3), 0, True),
@@ -1084,29 +1093,32 @@ class XgboostLocalTest(SparkTestCase):
                 min_child_weight=0.0,
                 reg_alpha=0,
                 reg_lambda=0,
+                tree_method=tree_method,
                 validation_indicator_col="val_col",
             )
             model = classifier.fit(df_train)
             pred_result = model.transform(df_train).collect()
             for row in pred_result:
-            self.assertEqual(row.prediction, 1.0)
+                assert row.prediction == 1.0
 
     def test_empty_partition(self):
         # raw_df.repartition(4) will result int severe data skew, actually,
         # there is no any data in reducer partition 1, reducer partition 2
         # see https://github.com/dmlc/xgboost/issues/8221
+        for tree_method in [
+            "hist",
+            "approx",
+        ]:  # pytest.mark conflict with python unittest
             raw_df = self.session.range(0, 100, 1, 50).withColumn(
-            "label", spark_sql_func.when(spark_sql_func.rand(1) > 0.5, 1).otherwise(0)
+                "label",
+                spark_sql_func.when(spark_sql_func.rand(1) > 0.5, 1).otherwise(0),
             )
             vector_assembler = (
                 VectorAssembler().setInputCols(["id"]).setOutputCol("features")
             )
             data_trans = vector_assembler.setHandleInvalid("keep").transform(raw_df)
-        data_trans.show(100)
 
-        classifier = SparkXGBClassifier(
+            classifier = SparkXGBClassifier(num_workers=4, tree_method=tree_method)
-            num_workers=4,
-        )
             classifier.fit(data_trans)
 
     def test_early_stop_param_validation(self):