#
# Copyright (c) 2021, NVIDIA CORPORATION.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import re
import numpy
from dask.dataframe.utils import meta_nonempty
from merlin.core.dispatch import DataFrameType, annotate
from merlin.schema import Schema
from .operator import ColumnSelector, Operator
[docs]class Groupby(Operator):
"""Groupby Transformation
Locally transform each partition of a Dataset with one or
more groupby aggregations.
WARNING: This transformation does NOT move data between
partitions. Please make sure that the target Dataset object
is already shuffled by ``groupby_cols``, otherwise the
output may be incorrect. See: ``Dataset.shuffle_by_keys``.
Example usage::
groupby_cols = ['user_id', 'session_id']
dataset = dataset.shuffle_by_keys(keys=groupby_cols)
groupby_features = [
'user_id', 'session_id', 'month', 'prod_id',
] >> ops.Groupby(
groupby_cols=groupby_cols,
sort_cols=['month'],
aggs={
'prod_id': 'list',
'month': ['first', 'last'],
},
)
processor = nvtabular.Workflow(groupby_features)
workflow.fit(dataset)
dataset_transformed = workflow.transform(dataset)
Parameters
-----------
groupby_cols : str or list of str
The column names to be used as groupby keys.
WARNING: Ensure the dataset was partitioned by those
groupby keys (see above for an example).
sort_cols : str or list of str
Columns to be used to sort each partition before
groupby aggregation is performed. If this argument
is not specified, the results will not be sorted.
aggs : dict, list or str
Groupby aggregations to perform. Supported list-based
aggregations include "list", "first" & "last". Most
conventional aggregations supported by Pandas/cuDF are
also allowed (e.g. "sum", "count", "max", "mean", etc.).
name_sep : str
String separator to use for new column names.
"""
def __init__(
self, groupby_cols=None, sort_cols=None, aggs="list", name_sep="_", ascending=True
):
self.groupby_cols = groupby_cols
self.sort_cols = sort_cols or []
if isinstance(self.groupby_cols, str):
self.groupby_cols = [self.groupby_cols]
if isinstance(self.sort_cols, str):
self.sort_cols = [self.sort_cols]
self.ascending = ascending
# Split aggregations into "conventional" aggregations
# and "list-based" aggregations. After this block,
# we will have a dictionary for each of these cases.
# We use the "__all__" key to specify aggregations
# that will be performed on all (non-key) columns.
self.list_aggs, self.conv_aggs = {}, {}
if isinstance(aggs, str):
aggs = {"__all__": [aggs]}
elif isinstance(aggs, list):
aggs = {"__all__": aggs}
for col, v in aggs.items():
_aggs = v if isinstance(v, list) else [v]
_conv_aggs, _list_aggs = set(), set()
for _agg in _aggs:
if is_list_agg(_agg):
_list_aggs.add("list" if _agg == list else _agg)
_conv_aggs.add(list)
else:
_conv_aggs.add(_agg)
if _conv_aggs:
self.conv_aggs[col] = list(_conv_aggs)
if _list_aggs:
self.list_aggs[col] = list(_list_aggs)
self.name_sep = name_sep
super().__init__()
transform.__doc__ = Operator.transform.__doc__
[docs] def compute_output_schema(
self, input_schema: Schema, col_selector: ColumnSelector, prev_output_schema: Schema = None
) -> Schema:
if not col_selector and hasattr(self, "target"):
col_selector = (
ColumnSelector(self.target) if isinstance(self.target, list) else self.target
)
return super().compute_output_schema(input_schema, col_selector, prev_output_schema)
[docs] def column_mapping(self, col_selector):
column_mapping = {}
for groupby_col in self.groupby_cols:
if groupby_col in col_selector.names:
column_mapping[groupby_col] = [groupby_col]
_list_aggs, _conv_aggs = _get_agg_dicts(
self.groupby_cols, self.list_aggs, self.conv_aggs, col_selector
)
for input_col_name, aggs in _list_aggs.items():
output_col_names = _columns_out_from_aggs(
{input_col_name: aggs}, name_sep=self.name_sep
)
for output_col_name in output_col_names:
column_mapping[output_col_name] = [input_col_name]
for input_col_name, aggs in _conv_aggs.items():
output_col_names = _columns_out_from_aggs(
{input_col_name: aggs}, name_sep=self.name_sep
)
for output_col_name in output_col_names:
column_mapping[output_col_name] = [input_col_name]
return column_mapping
def _compute_dtype(self, col_schema, input_schema):
col_schema = super()._compute_dtype(col_schema, input_schema)
dtype = col_schema.dtype
is_list = col_schema.is_list
dtypes = {
"count": numpy.int32,
"nunique": numpy.int32,
"mean": numpy.float32,
"var": numpy.float32,
"std": numpy.float32,
"median": numpy.float32,
}
is_lists = {"list": True}
for col_name in input_schema.column_names:
combined_aggs = _aggs_for_column(col_name, self.conv_aggs)
combined_aggs += _aggs_for_column(col_name, self.list_aggs)
for agg in combined_aggs:
if col_schema.name.endswith(f"{self.name_sep}{agg}"):
dtype = dtypes.get(agg, dtype)
is_list = is_lists.get(agg, is_list)
break
return col_schema.with_dtype(dtype, is_list=is_list, is_ragged=is_list)
def _aggs_for_column(col_name, agg_dict):
return agg_dict.get(col_name, []) + agg_dict.get("__all__", [])
def _columns_out_from_aggs(aggs, name_sep="_"):
# Helper function for `output_column_names`
_agg_cols = []
for k, v in aggs.items():
for _v in v:
if isinstance(_v, str):
_agg_cols.append(name_sep.join([k, _v]))
return _agg_cols
def _apply_aggs(_df, groupby_cols, _list_aggs, _conv_aggs, name_sep="_", ascending=True):
# Apply conventional aggs
_columns = list(set(groupby_cols) | set(_conv_aggs) | set(_list_aggs))
df = _df[_columns].groupby(groupby_cols).agg(_conv_aggs).reset_index()
df.columns = [
name_sep.join([n for n in name if n != ""]) for name in df.columns.to_flat_index()
]
# Handle custom aggs (e.g. "first" and "last")
for col, aggs in _list_aggs.items():
for _agg in aggs:
if is_list_agg(_agg, custom=True):
df[f"{col}{name_sep}{_agg}"] = _first_or_last(
df[f"{col}{name_sep}list"], _agg, ascending=ascending
)
if "list" not in aggs:
df.drop(columns=[col + f"{name_sep}list"], inplace=True)
for col in df.columns:
if re.search(f"{name_sep}(count|nunique)", col):
df[col] = df[col].astype(numpy.int32)
elif re.search(f"{name_sep}(mean|median|std|var)", col):
df[col] = df[col].astype(numpy.float32)
return df
def _get_agg_dicts(groupby_cols, list_aggs, conv_aggs, columns):
# Get updated aggregation dicts. This should map "__all__"
# to specific columns, and remove elements that are not
# in `columns`.
_allowed_cols = [c for c in columns.names if c not in groupby_cols]
_list_aggs = _ensure_agg_dict(list_aggs, _allowed_cols)
_conv_aggs = _ensure_agg_dict(conv_aggs, _allowed_cols)
return _list_aggs, _conv_aggs
def _ensure_agg_dict(_aggs, _allowed_cols):
# Make sure aggregation dict has legal keys
if "__all__" in _aggs:
return {col: _aggs["__all__"] for col in _allowed_cols}
else:
return {k: v for k, v in _aggs.items() if k in _allowed_cols}
def is_list_agg(agg, custom=False):
# check if `agg` is a supported list aggregation
if custom:
return agg in ("first", "last")
else:
return agg in ("list", list, "first", "last")
def _first_or_last(x, kind, ascending=True):
# Redirect to _first or _last
if kind == "first" and ascending:
return _first(x)
elif kind == "last" and not ascending:
return _first(x)
else:
return _last(x)
def _first(x):
# Convert each element of a list column to be the first
# item in the list
if hasattr(x, "list"):
# cuDF-specific behavior
offsets = x.list._column.offsets
elements = x.list._column.elements
return elements[offsets[:-1]]
else:
# cpu/pandas
return x.apply(lambda y: y[0])
def _last(x):
# Convert each element of a list column to be the last
# item in the list
if hasattr(x, "list"):
# cuDF-specific behavior
offsets = x.list._column.offsets
elements = x.list._column.elements
return elements[offsets[1:].values - 1]
else:
# cpu/pandas
return x.apply(lambda y: y[-1])