Backports v0.14.1 (#3055)

* Refactor tests for `ev.aggregations` (#3038)

* Fix edge cases in metric computation (#3037)

* Add plt.show() to README (#3043)

* Rotbaum: Add item-id to forecast. (#3049)

* Fix mypy checks (#3052)

---------

Co-authored-by: Jasper <schjaspe@amazon.de>

README.md

@@ -68,6 +68,7 @@ plt.plot(df["1954":], color="black")
 for forecast in forecasts:
  forecast.plot()
 plt.legend(["True values"], loc="upper left", fontsize="xx-large")
+plt.show()
 ```
 
 ![[train-test]](https://ts.gluon.ai/static/README/forecasts.png)

src/gluonts/ev/aggregations.py

@@ -50,7 +50,7 @@ class Sum(Aggregation):
  def step(self, values: np.ndarray) -> None:
  assert self.axis is None or isinstance(self.axis, tuple)
 
- summed_values = np.ma.sum(values, axis=self.axis)
+ summed_values = np.nansum(values, axis=self.axis)
 
  if self.axis is None or 0 in self.axis:
  if self.partial_result is None:
@@ -66,9 +66,11 @@ def get(self) -> np.ndarray:
  assert self.axis is None or isinstance(self.axis, tuple)
 
  if self.axis is None or 0 in self.axis:
- return np.ma.copy(self.partial_result)
+ assert isinstance(self.partial_result, np.ndarray)
+ return np.copy(self.partial_result)
 
- return np.ma.concatenate(self.partial_result)
+ assert isinstance(self.partial_result, list)
+ return np.concatenate(self.partial_result)
 
 
 @dataclass
@@ -107,7 +109,7 @@ def step(self, values: np.ndarray) -> None:
  if self.partial_result is None:
  self.partial_result = []
 
- mean_values = np.ma.mean(values, axis=self.axis)
+ mean_values = np.nanmean(values, axis=self.axis)
  assert isinstance(self.partial_result, list)
  self.partial_result.append(mean_values)
 
@@ -118,4 +120,5 @@ def get(self) -> np.ndarray:
  assert isinstance(self.partial_result, np.ndarray)
  return self.partial_result / self.n
 
- return np.ma.concatenate(self.partial_result)
+ assert isinstance(self.partial_result, list)
+ return np.concatenate(self.partial_result)

src/gluonts/ev/metrics.py

@@ -492,7 +492,7 @@ def mean(**quantile_losses: np.ndarray) -> np.ndarray:
  [quantile_loss for quantile_loss in quantile_losses.values()],
  axis=0,
  )
- return np.ma.mean(stacked_quantile_losses, axis=0)
+ return np.mean(stacked_quantile_losses, axis=0)
 
  def __call__(self, axis: Optional[int] = None) -> DerivedMetric:
  return DerivedMetric(
@@ -515,7 +515,7 @@ def mean(**quantile_losses: np.ndarray) -> np.ndarray:
  [quantile_loss for quantile_loss in quantile_losses.values()],
  axis=0,
  )
- return np.ma.mean(stacked_quantile_losses, axis=0)
+ return np.mean(stacked_quantile_losses, axis=0)
 
  def __call__(self, axis: Optional[int] = None) -> DerivedMetric:
  return DerivedMetric(
@@ -538,7 +538,7 @@ def mean(**quantile_losses: np.ndarray) -> np.ndarray:
  [quantile_loss for quantile_loss in quantile_losses.values()],
  axis=0,
  )
- return np.ma.mean(stacked_quantile_losses, axis=0)
+ return np.mean(stacked_quantile_losses, axis=0)
 
  def __call__(self, axis: Optional[int] = None) -> DerivedMetric:
  return DerivedMetric(
@@ -565,7 +565,7 @@ def mean(
  [np.abs(coverages[f"coverage[{q}]"] - q) for q in quantile_levels],
  axis=0,
  )
- return np.ma.mean(intermediate_result, axis=0)
+ return np.mean(intermediate_result, axis=0)
 
  def __call__(self, axis: Optional[int] = None) -> DerivedMetric:
  return DerivedMetric(

src/gluonts/ext/rotbaum/_predictor.py

@@ -50,12 +50,13 @@ def __init__(
  featurized_data: List,
  start_date: pd.Period,
  prediction_length: int,
+ item_id: Optional[str] = None,
  ):
  self.models = models
  self.featurized_data = featurized_data
  self.start_date = start_date
  self.prediction_length = prediction_length
- self.item_id = None
+ self.item_id = item_id
  self.lead_time = None
 
  def quantile(self, q: float) -> np.ndarray: # type: ignore
@@ -336,6 +337,7 @@ def predict( # type: ignore
  [featurized_data],
  start_date=forecast_start(ts),
  prediction_length=self.prediction_length,
+ item_id=ts.get("item_id"),
  )
 
  def explain(

src/gluonts/model/evaluation.py

@@ -203,7 +203,7 @@ def evaluate_forecasts(
  )
  if index0 is not None:
  index0_repeated = np.take(index0, indices=index_arrays[0], axis=0)
- index_arrays = (*zip(*index0_repeated), *index_arrays[1:])
+ index_arrays = (*zip(*index0_repeated), *index_arrays[1:]) # type: ignore
  index = pd.MultiIndex.from_arrays(index_arrays)
 
  flattened_metrics = valmap(np.ravel, metrics_values)

src/gluonts/mx/model/deepstate/issm.py

@@ -11,7 +11,7 @@
 # express or implied. See the License for the specific language governing
 # permissions and limitations under the License.
 
-from typing import List, Tuple
+from typing import List, Sequence, Tuple
 
 from pandas.tseries.frequencies import to_offset
 
@@ -31,7 +31,7 @@
 )
 
 
-def _make_block_diagonal(blocks: List[Tensor]) -> Tensor:
+def _make_block_diagonal(blocks: Sequence[Tensor]) -> Tensor:
  assert (
  len(blocks) > 0
  ), "You need at least one tensor to make a block-diagonal tensor"

test/ev/test_aggregations.py

@@ -18,84 +18,119 @@
 from gluonts.ev.aggregations import Mean, Sum
 from gluonts.itertools import power_set
 
-VALUE_STREAM = [
- [
- np.full((3, 5), np.nan),
- np.full((3, 5), np.nan),
- np.full((3, 5), np.nan),
- ],
- [
- np.array([[0, np.nan], [0, 0]]),
- np.array([[0, 5], [-5, np.nan]]),
- ],
- [
- np.full(shape=(3, 3), fill_value=1),
- np.full(shape=(1, 3), fill_value=4),
- ],
-]
-
-SUM_RES_AXIS_NONE = [
- 0,
- 0,
- 21,
-]
-
-SUM_RES_AXIS_0 = [
- np.zeros(5),
- np.array([-5, 5]),
- np.array([7, 7, 7]),
-]
-SUM_RES_AXIS_1 = [
- np.zeros(9),
- np.array([0, 0, 5, -5]),
- np.array([3, 3, 3, 12]),
-]
-
-
-MEAN_RES_AXIS_NONE = [
- np.nan,
- 0,
- 1.75,
-]
-
-MEAN_RES_AXIS_0 = [
- np.full(5, np.nan),
- np.array([-1.25, 2.5]),
- np.array([1.75, 1.75, 1.75]),
-]
-MEAN_RES_AXIS_1 = [
- np.full(9, np.nan),
- np.array([0, 0, 2.5, -5]),
- np.array([1, 1, 1, 4]),
-]
-
 
 @pytest.mark.parametrize(
  "value_stream, res_axis_none, res_axis_0, res_axis_1",
- zip(VALUE_STREAM, SUM_RES_AXIS_NONE, SUM_RES_AXIS_0, SUM_RES_AXIS_1),
+ [
+ (
+ [
+ np.full((3, 5), 0.0),
+ np.full((3, 5), 0.0),
+ np.full((3, 5), 0.0),
+ ],
+ 0.0,
+ np.zeros(5),
+ np.zeros(9),
+ ),
+ (
+ np.ma.masked_invalid(
+ [
+ np.full((3, 5), np.nan),
+ np.full((3, 5), np.nan),
+ np.full((3, 5), np.nan),
+ ]
+ ),
+ 0,
+ np.zeros(5),
+ np.zeros(9),
+ ),
+ (
+ np.ma.masked_invalid(
+ [
+ np.array([[0, np.nan], [0, 0]]),
+ np.array([[0, 5], [-5, np.nan]]),
+ ]
+ ),
+ 0,
+ np.array([-5, 5]),
+ np.array([0, 0, 5, -5]),
+ ),
+ (
+ [
+ np.full(shape=(3, 3), fill_value=1),
+ np.full(shape=(1, 3), fill_value=4),
+ ],
+ 21,
+ np.array([7, 7, 7]),
+ np.array([3, 3, 3, 12]),
+ ),
+ ],
 )
 def test_Sum(value_stream, res_axis_none, res_axis_0, res_axis_1):
  for axis, expected_result in zip(
  [None, 0, 1], [res_axis_none, res_axis_0, res_axis_1]
  ):
  sum = Sum(axis=axis)
  for values in value_stream:
- sum.step(np.ma.masked_invalid(values))
+ sum.step(values)
 
  np.testing.assert_almost_equal(sum.get(), expected_result)
 
 
 @pytest.mark.parametrize(
  "value_stream, res_axis_none, res_axis_0, res_axis_1",
- zip(VALUE_STREAM, MEAN_RES_AXIS_NONE, MEAN_RES_AXIS_0, MEAN_RES_AXIS_1),
+ [
+ (
+ [
+ np.full((3, 5), 0.0),
+ np.full((3, 5), 0.0),
+ np.full((3, 5), 0.0),
+ ],
+ 0.0,
+ np.zeros(5),
+ np.zeros(9),
+ ),
+ (
+ np.ma.masked_invalid(
+ [
+ np.full((3, 5), np.nan),
+ np.full((3, 5), np.nan),
+ np.full((3, 5), np.nan),
+ ]
+ ),
+ np.nan,
+ np.full(5, np.nan),
+ np.full(9, np.nan),
+ ),
+ (
+ np.ma.masked_invalid(
+ [
+ np.array([[0, np.nan], [0, 0]]),
+ np.array([[0, 5], [-5, np.nan]]),
+ ]
+ ),
+ 0,
+ np.array([-1.25, 2.5]),
+ np.array([0, 0, 2.5, -5]),
+ ),
+ (
+ [
+ np.full(shape=(3, 3), fill_value=1),
+ np.full(shape=(1, 3), fill_value=4),
+ ],
+ 1.75,
+ np.array([1.75, 1.75, 1.75]),
+ np.array([1, 1, 1, 4]),
+ ),
+ ],
 )
 def test_Mean(value_stream, res_axis_none, res_axis_0, res_axis_1):
  for axis, expected_result in zip(
  [None, 0, 1], [res_axis_none, res_axis_0, res_axis_1]
  ):
  mean = Mean(axis=axis)
  for values in value_stream:
- mean.step(np.ma.masked_invalid(values))
+ mean.step(values)
 
  np.testing.assert_almost_equal(mean.get(), expected_result)