[timeseries] Fix pandas groupby bug + GluonTS index bug #2420
Conversation
shchur
added
module: timeseries
|
Job PR-2420-28d8f49 is done. |
|
Job PR-2420-9523ddf is done. |
|
Job PR-2420-35c3780 is done. |
|
Job PR-2420-e992f1b is done. |
| @@ -76,7 +77,7 @@ def __iter__(self) -> Iterator[Dict[str, Any]]: | |||
| df = self.target_df.loc[item_id] | |||
| time_series = { | |||
| FieldName.ITEM_ID: item_id, | |||
| FieldName.TARGET: df.squeeze().to_numpy(dtype=self.float_dtype), | |||
There was a problem hiding this comment.
Is this a performance related change?
There was a problem hiding this comment.
Current code may lead to a bug. If df has length 1, then df.squeeze() will be a float instead of pd.Series, and calling to_numpy on this float throws an exception.
There is no clean way to convert a DataFrame to a Series (squeeze is the only function recommended for this purpose), so we rather do the flattening into 1D after converting to numpy with ravel.
Description of changes:
sort=False(related to BUG: df.groupby(sort=False) sorts multi-index-frames pandas-dev/pandas#17537)"2022-01-01 12:00","2022-01-02 12:00"(frequency is daily), GluonTS will produce forecasts with timestamps"2022-01-01 00:00","2022-01-02 00:00", which results in an AssertionError inside theTimeSeriesEvaluatorhereBy submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.