Merge pull request #1301 from thomcom/fea-ext-multiindex

Add MultiIndex support for Dataframes and Series

Author: thomcom

CHANGELOG.md

@@ -25,6 +25,7 @@
 - PR #1466 Add GPU-accelerated ORC Reader
 - PR #1565 Add build script for nightly doc builds
 - PR #1508 Add Series isna, isnull, and notna
+- PR #1301 MultiIndex support
 
 ## Improvements
 

ci/local/build.sh

@@ -256,6 +256,7 @@ gdf_error gdf_div_f64(gdf_column *lhs, gdf_column *rhs, gdf_column *output) {
 gdf_error F##_generic(gdf_column *lhs, gdf_column *rhs, gdf_column *output) { \
     switch ( lhs->dtype ) {                                                   \
     case GDF_INT8:      return F##_i8(lhs, rhs, output);                      \
+    case GDF_STRING_CATEGORY:                                                 \
     case GDF_INT32:     return F##_i32(lhs, rhs, output);                     \
     case GDF_INT64:     return F##_i64(lhs, rhs, output);                     \
     case GDF_FLOAT32:   return F##_f32(lhs, rhs, output);                     \

cpp/src/binary/binary_ops.cu

@@ -1,9 +1,9 @@
-# Copyright (c) 2018, NVIDIA CORPORATION.
+# Copyright (c) 2018-2019, NVIDIA CORPORATION.
 
 from cudf import dataframe
 from cudf import datasets
 from cudf.dataframe import DataFrame, from_pandas, merge
-from cudf.dataframe import Index
+from cudf.dataframe import Index, MultiIndex
 from cudf.dataframe import Series
 from cudf.multi import concat
 from cudf.io import (read_csv, read_parquet, read_feather, read_json,

python/cudf/__init__.py

@@ -334,12 +334,10 @@ def agg(groupby_class, args):
                 sort_results=sort_results
             )
             add_col_values = False  # we only want to add them once
-        # TODO: Do multindex here
-        if(groupby_class._as_index) and 1 == len(groupby_class._by):
-            idx = index.as_index(result[groupby_class._by[0]])
-            idx.name = groupby_class._by[0]
-            result = result.set_index(idx)
-            result.drop_column(idx.name)
+        if(groupby_class._as_index):
+            result = groupby_class.apply_multiindex_or_single_index(result)
+        if use_prefix:
+            result = groupby_class.apply_multicolumn(result, args)
     elif isinstance(args, collections.abc.Mapping):
         if (len(args.keys()) == 1):
             if(len(list(args.values())[0]) == 1):
@@ -377,15 +375,13 @@ def agg(groupby_class, args):
                     sort_results=sort_results
                 )
             add_col_values = False  # we only want to add them once
-        # TODO: Do multindex here
-        if(groupby_class._as_index) and 1 == len(groupby_class._by):
-            idx = index.as_index(result[groupby_class._by[0]])
-            idx.name = groupby_class._by[0]
-            result = result.set_index(idx)
-            result.drop_column(idx.name)
+        if groupby_class._as_index:
+            result = groupby_class.apply_multiindex_or_single_index(result)
+        if use_prefix:
+            result = groupby_class.apply_multicolumn_mapped(result, args)
     else:
         result = groupby_class.agg([args])
-        
+
     free(ctx)
 
     nvtx_range_pop()
@@ -431,18 +427,13 @@ def _apply_basic_agg(groupby_class, agg_type, sort_results=False):
         else:
             idx.name = groupby_class._by[0]
         result_series = result_series.set_index(idx)
+        if groupby_class._as_index:
+            result = groupby_class.apply_multiindex_or_single_index(result)
+            result_series.index = result.index
         return result_series
 
-    # TODO: Do MultiIndex here
-    if(groupby_class._as_index):
-        idx = index.as_index(result[groupby_class._by[0]])
-        idx.name = groupby_class._by[0]
-        result.drop_column(idx.name)
-        if groupby_class.level == 0:
-            idx.name = groupby_class._original_index_name
-        else:
-            idx.name = groupby_class._by[0]
-        result = result.set_index(idx)
+    if groupby_class._as_index:
+        result = groupby_class.apply_multiindex_or_single_index(result)
 
     nvtx_range_pop()
 

python/cudf/bindings/groupby.pyx

@@ -1,9 +1,12 @@
+# Copyright (c) 2018-2019, NVIDIA CORPORATION.
+
 from cudf.dataframe import (buffer, dataframe, series,
     index, numerical, datetime, categorical, string)
 
 from cudf.dataframe.dataframe import DataFrame, from_pandas, merge
 from cudf.dataframe.index import (Index, GenericIndex,
     RangeIndex, DatetimeIndex, CategoricalIndex)
+from cudf.dataframe.multiindex import MultiIndex
 from cudf.dataframe.series import Series
 from cudf.dataframe.buffer import Buffer
 from cudf.dataframe.numerical import NumericalColumn

python/cudf/dataframe/__init__.py

@@ -1,4 +1,4 @@
-# Copyright (c) 2018, NVIDIA CORPORATION.
+# Copyright (c) 2018-2019, NVIDIA CORPORATION.
 
 from __future__ import print_function, division
 
@@ -29,6 +29,7 @@
 
 from librmm_cffi import librmm as rmm
 
+import cudf
 from cudf import formatting
 from cudf.utils import cudautils, queryutils, applyutils, utils, ioutils
 from cudf.dataframe.index import as_index, Index, RangeIndex
@@ -224,10 +225,14 @@ def __getitem__(self, arg):
         >>> print(df[[True, False, True, False]]) # mask the entire dataframe,
         # returning the rows specified in the boolean mask
         """
+        if isinstance(self.columns, cudf.dataframe.multiindex.MultiIndex) and\
+           isinstance(arg, tuple):
+            return self.columns._get_column_major(self, arg)
         if isinstance(arg, str) or isinstance(arg, numbers.Integral) or \
            isinstance(arg, tuple):
             s = self._cols[arg]
             s.name = arg
+            s.index = self.index
             return s
         elif isinstance(arg, slice):
             df = DataFrame()
@@ -247,7 +252,7 @@ def __getitem__(self, arg):
                 index = self.index.take(selinds.to_gpu_array())
                 for col in self._cols:
                     df[col] = Series(self._cols[col][arg], index=index)
-                df.set_index(index)
+                df = df.set_index(index)
             else:
                 for col in arg:
                     df[col] = self[col]
@@ -272,7 +277,6 @@ def mask(self, other):
     def __setitem__(self, name, col):
         """Add/set column by *name or DataFrame*
         """
-        # div[div < 0] = 0
         if isinstance(name, DataFrame):
             for col_name in self._cols:
                 mask = name[col_name]
@@ -399,6 +403,11 @@ def to_string(self, nrows=NOTSET, ncols=NOTSET):
         >>> df.to_string()
         '   key   val\\n0    0  10.0\\n1    1  11.0\\n2    2  12.0'
         """
+        if isinstance(self.index, cudf.dataframe.multiindex.MultiIndex) or\
+           isinstance(self.columns, cudf.dataframe.multiindex.MultiIndex):
+            raise TypeError("You're trying to print a DataFrame that contains "
+                            "a MultiIndex. Print this dataframe with "
+                            ".to_pandas()")
         if nrows is NOTSET:
             nrows = settings.formatting.get('nrows')
         if ncols is NOTSET:
@@ -420,9 +429,12 @@ def to_string(self, nrows=NOTSET, ncols=NOTSET):
         # Prepare cells
         cols = OrderedDict()
         dtypes = OrderedDict()
-        use_cols = list(self.columns[:ncols - 1])
-        if ncols > 0:
-            use_cols.append(self.columns[-1])
+        if hasattr(self, 'multi_cols'):
+            use_cols = list(range(len(self.columns)))
+        else:
+            use_cols = list(self.columns[:ncols - 1])
+            if ncols > 0:
+                use_cols.append(self.columns[-1])
 
         for h in use_cols:
             cols[h] = self[h].values_to_string(nrows=nrows)
@@ -664,19 +676,41 @@ def iloc(self):
     def columns(self):
         """Returns a tuple of columns
         """
-        return pd.Index(self._cols)
+        if hasattr(self, 'multi_cols'):
+            return self.multi_cols
+        else:
+            return pd.Index(self._cols)
 
     @columns.setter
     def columns(self, columns):
+        if isinstance(columns, Index):
+            if len(columns) != len(self.columns):
+                msg = f"Length mismatch: Expected axis has %d elements, "\
+                        "new values have %d elements"\
+                        % (len(self.columns), len(columns))
+                raise ValueError(msg)
+            """
+            new_names = []
+            for idx, name in enumerate(columns):
+                new_names.append(name)
+            self._rename_columns(new_names)
+            """
+            self.multi_cols = columns
+        else:
+            if hasattr(self, 'multi_cols'):
+                delattr(self, 'multi_cols')
+            self._rename_columns(columns)
+
+    def _rename_columns(self, new_names):
         old_cols = list(self._cols.keys())
         l_old_cols = len(old_cols)
-        l_new_cols = len(columns)
+        l_new_cols = len(new_names)
         if l_new_cols != l_old_cols:
             msg = f'Length of new column names: {l_new_cols} does not ' \
                   'match length of previous column names: {l_old_cols}'
             raise ValueError(msg)
 
-        mapper = dict(zip(old_cols, columns))
+        mapper = dict(zip(old_cols, new_names))
         self.rename(mapper=mapper, inplace=True)
 
     @property
@@ -687,12 +721,26 @@ def index(self):
 
     @index.setter
     def index(self, _index):
+        if isinstance(_index, cudf.dataframe.multiindex.MultiIndex):
+            if len(_index) != len(self[self.columns[0]]):
+                msg = f"Length mismatch: Expected axis has "\
+                       "%d elements, new values "\
+                       "have %d elements"\
+                       % (len(self[self.columns[0]]), len(_index))
+                raise ValueError(msg)
+            self._index = _index
+            for k in self.columns:
+                self[k].index = _index
+            return
+
         new_length = len(_index)
         old_length = len(self._index)
 
         if new_length != old_length:
-            msg = f'Length mismatch: Expected index has {old_length}' \
-                    ' elements, new values have {new_length} elements'
+            msg = f"Length mismatch: Expected axis has "\
+                   "%d elements, new values "\
+                   "have %d elements"\
+                   % (old_length, new_length)
             raise ValueError(msg)
 
         # try to build an index from generic _index
@@ -906,8 +954,8 @@ def drop(self, labels, axis=None):
         if axis == 0:
             raise NotImplementedError("Can only drop columns, not rows")
 
-        columns = [labels] if isinstance(labels, str) else list(labels)
-
+        columns = [labels] if isinstance(
+                labels, (str, numbers.Number)) else list(labels)
         outdf = self.copy()
         for c in columns:
             outdf._drop_column(c)
@@ -2240,6 +2288,13 @@ def to_pandas(self):
         out = pd.DataFrame(index=index)
         for c, x in self._cols.items():
             out[c] = x.to_pandas(index=index)
+        if isinstance(self.columns, Index):
+            out.columns = self.columns
+            if isinstance(self.columns, cudf.dataframe.multiindex.MultiIndex):
+                if self.columns.names is not None:
+                    out.columns.names = self.columns.names
+            else:
+                out.columns.name = self.columns.name
         return out
 
     @classmethod
@@ -2269,7 +2324,12 @@ def from_pandas(cls, dataframe, nan_as_null=True):
             vals = dataframe[colk].values
             df[colk] = Series(vals, nan_as_null=nan_as_null)
         # Set index
-        return df.set_index(dataframe.index)
+        if isinstance(dataframe.index, pd.MultiIndex):
+            import cudf
+            index = cudf.from_pandas(dataframe.index)
+        else:
+            index = dataframe.index
+        return df.set_index(index)
 
     def to_arrow(self, preserve_index=True):
         """
@@ -2696,6 +2756,13 @@ def __getitem__(self, arg):
         row_slice = None
         row_label = None
 
+        if isinstance(self._df.index, cudf.dataframe.multiindex.MultiIndex)\
+                and isinstance(arg, tuple):  # noqa: E501
+            # Explicitly ONLY support tuple indexes into MultiIndex.
+            # Pandas allows non tuple indices and warns "results may be
+            # undefined."
+            return self._df._index._get_row_major(self._df, arg)
+
         if isinstance(arg, int):
             if arg < 0 or arg >= len(self._df):
                 raise IndexError("label scalar %s is out of bound" % arg)
@@ -2785,7 +2852,9 @@ def __setitem__(self, key, value):
 
 def from_pandas(obj):
     """
-    Convert a Pandas DataFrame or Series object into the cudf equivalent
+    Convert certain Pandas objects into the cudf equivalent.
+
+    Supports DataFrame, Series, or MultiIndex.
 
     Raises
     ------
@@ -2804,9 +2873,12 @@ def from_pandas(obj):
         return DataFrame.from_pandas(obj)
     elif isinstance(obj, pd.Series):
         return Series.from_pandas(obj)
+    elif isinstance(obj, pd.MultiIndex):
+        return cudf.dataframe.multiindex.MultiIndex.from_pandas(obj)
     else:
         raise TypeError(
-            "from_pandas only accepts Pandas Dataframes and Series objects. "
+            "from_pandas only accepts Pandas Dataframes, Series, and "
+            "MultiIndex objects. "
             "Got %s" % type(obj)
         )