Encode pandas categorical data automatically.

python-package/xgboost/data.py

@@ -220,12 +220,13 @@ def _is_modin_df(data):
 
 def _transform_pandas_df(
     data,
-    enable_categorical,
+    enable_categorical: bool,
     feature_names: Optional[List[str]] = None,
     feature_types: Optional[List[str]] = None,
     meta=None,
     meta_type=None,
 ):
+    import pandas as pd
     from pandas import MultiIndex, Int64Index, RangeIndex
     from pandas.api.types import is_sparse, is_categorical_dtype
 
@@ -238,11 +239,12 @@ def _transform_pandas_df(
             if dtype.name not in _pandas_dtype_mapper
         ]
 
-        msg = """DataFrame.dtypes for data must be int, float, bool or categorical.  When
-                categorical type is supplied, DMatrix parameter
-                `enable_categorical` must be set to `True`."""
+        msg = """DataFrame.dtypes for data must be int, float, bool or category.  When
+                categorical type is supplied, DMatrix parameter `enable_categorical` must
+                be set to `True`."""
         raise ValueError(msg + ', '.join(bad_fields))
 
+    # handle feature names
     if feature_names is None and meta is None:
         if isinstance(data.columns, MultiIndex):
             feature_names = [
@@ -253,28 +255,39 @@ def _transform_pandas_df(
         else:
             feature_names = data.columns.format()
 
+    # handle feature types
     if feature_types is None and meta is None:
         feature_types = []
-        for dtype in data_dtypes:
+        for i, dtype in enumerate(data_dtypes):
             if is_sparse(dtype):
-                feature_types.append(_pandas_dtype_mapper[
-                    dtype.subtype.name])
+                feature_types.append(_pandas_dtype_mapper[dtype.subtype.name])
             elif is_categorical_dtype(dtype) and enable_categorical:
                 feature_types.append(CAT_T)
             else:
                 feature_types.append(_pandas_dtype_mapper[dtype.name])
 
+    # handle categorical codes.
+    transformed = pd.DataFrame()
+    if enable_categorical:
+        for i, dtype in enumerate(data_dtypes):
+            if is_categorical_dtype(dtype):
+                transformed[data.columns[i]] = data[data.columns[i]].cat.codes
+            else:
+                transformed[data.columns[i]] = data[data.columns[i]]
+    else:
+        transformed = data
+
     if meta and len(data.columns) > 1:
         raise ValueError(
             'DataFrame for {meta} cannot have multiple columns'.format(
                 meta=meta)
         )
 
     dtype = meta_type if meta_type else np.float32
-    data = data.values
+    arr = transformed.values
     if meta_type:
-        data = data.astype(meta_type)
-    return data, feature_names, feature_types
+        arr = arr.astype(meta_type)
+    return arr, feature_names, feature_types
 
 
 def _from_pandas_df(

src/tree/updater_gpu_hist.cu

@@ -582,6 +582,8 @@ struct GPUHistMakerDevice {
 
     auto is_cat = candidate.split.is_cat;
     if (is_cat) {
+      CHECK_LT(candidate.split.fvalue, std::numeric_limits<bst_cat_t>::max())
+          << "Categorical feature value too large.";
       auto cat = common::AsCat(candidate.split.fvalue);
       std::vector<uint32_t> split_cats(LBitField32::ComputeStorageSize(std::max(cat+1, 1)), 0);
       LBitField32 cats_bits(split_cats);

tests/python/test_with_pandas.py

@@ -130,6 +130,17 @@ def test_pandas_categorical(self):
         m = xgb.DMatrix(X, y, enable_categorical=True)
         assert m.feature_types[0] == 'c'
 
+        X_0 = ["f", "o", "o"]
+        X_1 = [4, 3, 2]
+        X = pd.DataFrame({"feat_0": X_0, "feat_1": X_1})
+        X["feat_0"] = X["feat_0"].astype("category")
+        transformed, _, feature_types = xgb.data._transform_pandas_df(
+            X, enable_categorical=True
+        )
+
+        assert np.issubdtype(transformed[:, 0].dtype, np.integer)
+        assert transformed[:, 0].min() == 0
+
     def test_pandas_sparse(self):
         import pandas as pd
         rows = 100