Lorenzo Stella

Otherwise this appears to be making the input transform significantly simpler to define. Applying this to all estimators will be a good testbed for this proposal.

> I like the opt-in behaviour by using a decorator. It is also more explicit looking at the code. Eh, not sure it’s more explicit than ```python training_data = Map(schema,...

> Is there any reason to keep make_evaluation_predictions? Eventually no, I think. Right now it's a shortcut to get backetesting forecasts so you can plot them for example, instead of...

Minimal reproducer inspired by the referenced discussion ```python import numpy as np from gluonts.dataset.common import ListDataset from gluonts.mx.trainer import Trainer from gluonts.model.deepar import DeepAREstimator data = [] for i in...

Note that a similar example running the Torch version of the estimator fails gracefully ```python import numpy as np from gluonts.dataset.common import ListDataset from gluonts.torch.model.deepar import DeepAREstimator data = []...

"Gracefully" was an overstatement maybe, but at least training raises an exception (so no model is even produced). This happens since PyTorch performs bound checks on the parameters of the...

The problem shows up at prediction time in the case of MXNet models, yes. Maybe to understand this better, one could train on perfectly fine data (no `inf` values), get...

I also don't think there's any use for `inf` values in the data (unlike `nan`, which are used to indicate missing values), but maybe we can think about that later....

Looks like so. I suppose it works locally? If so, something really strange is happening in the test environment…

@zoolhasson could you take a look? It looks like `_pre_transform` is missing some arguments in the calls [here](https://github.com/awslabs/gluon-ts/blob/b47a602dbb2dce8326a27348e0ae8cd2d8ee46ee/src/gluonts/model/rotbaum/_preprocess.py#L434) and [here](https://github.com/awslabs/gluon-ts/blob/b47a602dbb2dce8326a27348e0ae8cd2d8ee46ee/src/gluonts/model/rotbaum/_preprocess.py#L449), but I'm not sure what the behavior is supposed to...