Hi,
I just installed version 0.5.1 via pip install --upgrade mxnet==1.6 gluonts and stumbled into an error using the OffsetSplitter.
The traceback is the following:
AttributeError Traceback (most recent call last)
<ipython-input-9-157d13f7a285> in <module>
1 splitter = OffsetSplitter(prediction_length=24, split_offset=24)
----> 2 train, test = splitter.split(items=whole_dataset)
~/opt/anaconda3/envs/tmp_dem/lib/python3.8/site-packages/gluonts/dataset/split/splitter.py in split(self, items)
195 split = TrainTestSplit()
196
--> 197 for item in map(TimeSeriesSlice.from_time_series_item, items):
198 train = self._train_slice(item)
199 test = self._trim_history(self._test_slice(item))
~/opt/anaconda3/envs/tmp_dem/lib/python3.8/site-packages/gluonts/dataset/split/splitter.py in from_time_series_item(cls, item, freq)
76 ) -> "TimeSeriesSlice":
77 if freq is None:
---> 78 freq = item.start.freq
79
80 index = pd.date_range(
AttributeError: 'dict' object has no attribute 'start'
The failing line was already fixed in master in this PR.
That got me wondering how the old implementation ended up in the release.
I think that the released version 0.5.1 is of a different branch than master, previous releases were of the branch master.
The tagged commit 22e7aa7 (link) is associated to a PR for the branch v0.5.x.
Is it possible to get the recent master fixes included in the new releases?
Using the latest master via pip install git+git://github.com/awslabs/gluon-ts.git#egg=gluon-ts on the other hand works like a charm and no errors are raised.
So the issue is that the latest release does not include recent master changes.
@MaximilianProll this seems like a missing backport, I'll try to write down a list of fixes that are needed in the 0.5.x branch
@MaximilianProll can you confirm that this is fixed in 0.5.2?
Wow, you guys are fast! 馃殌 Thank you, yes, version 0.5.2 now includes the relevant fixes for the OffsetSplitter.
Could this PR also be included in the recent releases or was that left out on purpose?
Good point! Added that to #1016, feel free to comment there if you spot anything else.
Closing this, thanks a lot for pointing this out.