Add LongDataset.

docs/tutorials/data_manipulation/pandasdataframes.md.template

@@ -74,12 +74,12 @@ df.head()
 ```
 
 After reading the data into a `pandas.DataFrame` we can easily convert
-it to `gluonts.dataset.pandas.PandasDataset` and train an estimator an get forecasts.
+it into a `Dataset` and train an estimator an get forecasts.
 
 ```python
-from gluonts.dataset.pandas import PandasDataset
+from gluonts.dataset.pandas import LongDataset
 
-ds = PandasDataset.from_long_dataframe(df, target="target", item_id="item_id")
+ds = LongDataset(df, item_id="item_id")
 ```
 
 ```python

src/gluonts/dataset/pandas.py

@@ -12,15 +12,19 @@
 # permissions and limitations under the License.
 
 from copy import deepcopy
-from dataclasses import dataclass, field
+from dataclasses import dataclass, field, InitVar
+from functools import partial
+from operator import methodcaller
 from typing import Any, cast, Dict, Iterator, List, Optional, Union
 
 import pandas as pd
 from pandas.core.indexes.datetimelike import DatetimeIndexOpsMixin
-from toolz import valmap
+from toolz import valmap, second
 
-from gluonts.dataset.common import DataEntry, ProcessDataEntry
-from gluonts.dataset.field_names import FieldName
+from gluonts.itertools import Map
+from .common import DataEntry, ProcessDataEntry, _as_period
+from .field_names import FieldName
+from .schema import Translator
 
 
 @dataclass
@@ -152,35 +156,6 @@ def __iter__(self) -> Iterator[DataEntry]:
     def __len__(self) -> int:
         return len(self._dataframes)
 
-    @classmethod
-    def from_long_dataframe(
-        cls, dataframe: pd.DataFrame, item_id: str, **kwargs
-    ) -> "PandasDataset":
-        """
-        Construct ``PandasDataset`` out of a long dataframe.
-        A long dataframe uses the long format for each variable. Target time
-        series values, for example, are stacked on top of each other rather
-        than side-by-side. The same is true for other dynamic or categorical
-        features.
-
-        Parameters
-        ----------
-        dataframe
-            pandas.DataFrame containing at least ``timestamp``, ``target`` and
-            ``item_id`` columns.
-        item_id
-            Name of the column that, when grouped by, gives the different time
-            series.
-        **kwargs
-            Additional arguments. Same as of PandasDataset class.
-
-        Returns
-        -------
-        PandasDataset
-            Gluonts dataset based on ``pandas.DataFrame``s.
-        """
-        return cls(dataframes=dict(list(dataframe.groupby(item_id))), **kwargs)
-
 
 def series_to_dataframe(
     series: Union[pd.Series, List[pd.Series], Dict[str, pd.Series]]
@@ -281,10 +256,12 @@ def prepare_prediction_data(
     Remove ``ignore_last_n_targets`` values from ``target`` and
     ``past_feat_dynamic_real``.  Works in univariate and multivariate case.
 
-    >>> prepare_prediction_data(
-    >>>    {"target": np.array([1., 2., 3., 4.])}, ignore_last_n_targets=2
-    >>> )
-    {'target': array([1., 2.])}
+        >>> import numpy as np
+        >>> prepare_prediction_data(
+        ...    {"target": np.array([1., 2., 3., 4.])}, ignore_last_n_targets=2
+        ... )
+        {'target': array([1., 2.])}
+
     """
     entry = deepcopy(dataentry)
     for fname in [FieldName.TARGET, FieldName.PAST_FEAT_DYNAMIC_REAL]:
@@ -298,12 +275,139 @@ def is_uniform(index: pd.PeriodIndex) -> bool:
     Check if ``index`` contains monotonically increasing periods, evenly spaced
     with frequency ``index.freq``.
 
-    >>> ts = ["2021-01-01 00:00", "2021-01-01 02:00", "2021-01-01 04:00"]
-    >>> is_uniform(pd.DatetimeIndex(ts).to_period("2H"))
-    True
-    >>> ts = ["2021-01-01 00:00", "2021-01-01 04:00"]
-    >>> is_uniform(pd.DatetimeIndex(ts).to_period("2H"))
-    False
+        >>> ts = ["2021-01-01 00:00", "2021-01-01 02:00", "2021-01-01 04:00"]
+        >>> is_uniform(pd.DatetimeIndex(ts).to_period("2H"))
+        True
+        >>> ts = ["2021-01-01 00:00", "2021-01-01 04:00"]
+        >>> is_uniform(pd.DatetimeIndex(ts).to_period("2H"))
+        False
+
     """
     other = pd.period_range(index[0], periods=len(index), freq=index.freq)
     return (other == index).all()
+
+
+def _column_as_start(dct: dict, column: str, freq, unchecked=False) -> dict:
+    ts = dct.pop(column)
+
+    if unchecked:
+        dct["start"] = _as_period(ts[0], freq=freq)
+    else:
+        idx = pd.PeriodIndex(ts, freq=freq)
+        assert is_uniform(idx)
+        dct["start"] = idx[0]
+
+    return dct
+
+
+def _drop_index(df, name):
+    index = df.index
+    df = df.reset_index(drop=True)
+    df[name] = index
+    return df
+
+
+@dataclass
+class LongDataset:
+    """Wrapper for ``pandas.DataFrame`` in long format.
+
+    Given a dataframe and an item identifier, this will yield a ``DataEntry``
+    for each item, by first calling ``dataframe.groupby(item_id)``.
+
+    A time-column is optional, but if `timestamp` is provided, `freq` needs to
+    be provided as well.
+
+    ``translate`` can be set to rename and stack columns on the fly.
+
+    Since the result of the ``groupby`` operation is not guaranteed to be in
+    right order, each group is sorted by default. If ``assume_sorted`` is set
+    to ``True``, sorting is skipped which improves performance. Note however,
+    that if the dataframe is unordered, this will lead to incorrect behaviour.
+
+    To improve performance further, ``unchecked`` can be set to ``True`` to
+    indicate that the timestamp column is uniform (see ``is_uniform``).
+    """
+
+    df: pd.DataFrame
+    item_id: Union[str, List[str]]
+    timestamp: Optional[str] = None
+    freq: Optional[str] = None
+
+    assume_sorted: bool = False
+    translate: InitVar[Optional[dict]] = None
+    unchecked: bool = False
+    translator: Optional[Translator] = field(default=None, init=False)
+    use_index: bool = field(init=False)
+
+    def __post_init__(self, translate):
+        self.use_index = False
+
+        if self.timestamp is None:
+            if isinstance(self.df.index, DatetimeIndexOpsMixin):
+                if self.freq is None:
+                    assert self.df.index.freq is not None
+                    self.freq = self.df.index.freq
+
+                self.use_index = True
+                self.timestamp = "__timestamp_index__"
+
+        elif self.freq is None:
+            raise ValueError(
+                "When providing `timestamp`, `freq` needs to be provided too."
+            )
+
+        if translate is not None:
+            self.translator = Translator.parse(translate, drop=True)
+        else:
+            self.translator = None
+
+    def _handle_item_id(self, dct):
+        """Ensure field "item_id" is a single string."""
+        if isinstance(self.item_id, list):
+            dct["item_id"] = ", ".join(
+                dct.pop(column)[0] for column in self.item_id
+            )
+        else:
+            dct["item_id"] = dct.pop(self.item_id)[0]
+
+        return dct
+
+    def __iter__(self):
+        groups = self.df.groupby(self.item_id)
+        # groups contains tuples of (item_id, df), but we only want df
+        dataset = Map(second, groups)
+
+        if self.use_index:
+            dataset = Map(
+                partial(_drop_index, name=self.timestamp),
+                dataset,
+            )
+
+        if not self.assume_sorted and self.timestamp is not None:
+            dataset = Map(
+                methodcaller("sort_values", by=self.timestamp), dataset
+            )
+
+        dataset = Map(methodcaller("to_dict", orient="list"), dataset)
+        dataset = Map(self._handle_item_id, dataset)
+
+        if self.translator is not None:
+            dataset = Map(self.translator, dataset)
+
+        # handle start field
+        if self.timestamp is not None:
+            dataset = Map(
+                partial(
+                    _column_as_start,
+                    column=self.timestamp,
+                    freq=self.freq,
+                    unchecked=self.unchecked,
+                ),
+                dataset,
+            )
+
+        yield from dataset
+        self.unchecked = True
+
+    def __len__(self):
+        return len(self.df.groupby(self.item_id))

src/gluonts/dataset/polars.py

@@ -0,0 +1,99 @@
+# Copyright 2018 Amazon.com, Inc. or its affiliates. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License").
+# You may not use this file except in compliance with the License.
+# A copy of the License is located at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# or in the "license" file accompanying this file. This file 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.
+
+from dataclasses import dataclass, field, InitVar
+from functools import partial
+from operator import methodcaller
+from typing import Union, List, Optional
+
+# import pandas as pd
+import polars as pl
+
+from gluonts.itertools import Map
+from .pandas import _column_as_start
+from .schema import Translator
+
+
+@dataclass
+class LongDataset:
+    df: pl.DataFrame
+    item_id: Union[str, List[str]]
+    timestamp: Optional[str] = None
+    freq: Optional[str] = None
+
+    assume_sorted: bool = False
+    translate: InitVar[Optional[dict]] = None
+    translator: Optional[Translator] = field(default=None, init=False)
+
+    use_partition: bool = False
+    unchecked: bool = False
+
+    def __post_init__(self, translate):
+        if (self.timestamp is None) != (self.freq is None):
+            raise ValueError(
+                "Either both `timestamp` and `freq` have to be "
+                "provided or neither."
+            )
+
+        if translate is not None:
+            self.translator = Translator.parse(translate, drop=True)
+        else:
+            self.translator = None
+
+    def _pop_item_id(self, dct):
+        if isinstance(self.item_id, list):
+            dct["item_id"] = ", ".join(
+                dct.pop(column)[0] for column in self.item_id
+            )
+        else:
+            dct["item_id"] = dct.pop(self.item_id)[0]
+
+        return dct
+
+    def __iter__(self):
+        if self.use_partition:
+            dataset = self.df.partition_by(self.item_id)
+        else:
+            dataset = self.df.groupby(self.item_id)
+
+        if not self.assume_sorted:
+            sort_by = [self.item_id]
+            if self.timestamp is not None:
+                sort_by.append(self.timestamp)
+                dataset = Map(methodcaller("sort", by=sort_by), dataset)
+
+        dataset = Map(methodcaller("to_dict", as_series=True), dataset)
+        dataset = Map(self._pop_item_id, dataset)
+
+        if self.translator is not None:
+            dataset = Map(self.translator, dataset)
+
+        if self.timestamp is not None:
+            dataset = Map(
+                partial(
+                    _column_as_start,
+                    column=self.timestamp,
+                    freq=self.freq,
+                    unchecked=self.unchecked,
+                ),
+                dataset,
+            )
+
+        yield from dataset
+
+        # we were successful to iterate once over the dataset
+        # so no more need to check more
+        self.unchecked = True
+
+    def __len__(self):
+        return len(self.df.groupby(self.item_id).count())

test/dataset/test_pandas.py

@@ -60,14 +60,16 @@ def long_dataframe():
 
 @pytest.fixture
 def long_dataset(long_dataframe):  # initialized with dict
-    return pandas.PandasDataset.from_long_dataframe(
-        dataframe=long_dataframe,
-        target="target",
+    return pandas.LongDataset(
+        long_dataframe,
         timestamp="time",
         item_id="item",
         freq="1H",
-        feat_dynamic_real=["dyn_real_1"],
-        feat_static_cat=["stat_cat_1"],
+        translate=dict(
+            target="target",
+            feat_dynamic_real=["dyn_real_1"],
+            feat_static_cat=["stat_cat_1"],
+        ),
     )
 
 
@@ -221,9 +223,8 @@ def test_long_csv_3M():
     )
 
     with io.StringIO(data) as fp:
-        ds = pandas.PandasDataset.from_long_dataframe(
+        ds = pandas.LongDataset(
             pd.read_csv(fp),
-            target="target",
             item_id="item_id",
             timestamp="timestamp",
             freq="3M",
@@ -232,9 +233,8 @@ def test_long_csv_3M():
             assert entry["start"].freqstr == "3M"
 
     with io.StringIO(data) as fp:
-        ds = pandas.PandasDataset.from_long_dataframe(
-            pd.read_csv(fp, index_col="timestamp"),
-            target="target",
+        ds = pandas.LongDataset(
+            pd.read_csv(fp, index_col="timestamp", parse_dates=True),
             item_id="item_id",
             freq="3M",
         )