SimpleFeedForward failing when input dataset has dynamic real features of different lengths

[StanislasGuinel]

Hi,
When I train a SimpleFeedForwardEstimator on a ListDataset with time series of different lengths and "feat_dynamic_real" fields, it fails.
I know SimpleFeedForward doesn't support dynamic real features, but in my use case, I train different models on the same dataset and some of the models use dynamic real features, it would be nice for SimpleFeedForward not to fail on such datasets.

Here is some code to reproduce the error:

from gluonts.dataset.common import ListDataset
from gluonts.model.deepar import DeepAREstimator
from gluonts.model.simple_feedforward import SimpleFeedForwardEstimator
from gluonts.trainer import Trainer
training_data = ListDataset(
    data_iter=[
        {"start": "2019-01-01 00:00:00", "target": [1.0, 2.0, 3.0, 4.0],
         "feat_dynamic_real": [[1.0, 2.0, 3.0, 4.0]]},
        {"start": "2019-01-01 00:00:00", "target": [1.0, 2.0, 3.0, 4.0],
         "feat_dynamic_real": [[1.0, 2.0, 3.0, 4.0]]},
    ],
    freq="D"
)
deepar_estimator = DeepAREstimator(freq="D", prediction_length=2, trainer=Trainer(epochs=1))
feed_forward_estimator = SimpleFeedForwardEstimator(freq="D", prediction_length=2, trainer=Trainer(epochs=1))
deepar_predictor = deepar_estimator.train(training_data)
feed_forward_predictor = feed_forward_estimator.train(training_data)

DeepAR works fine and here is the error I get on SimpleFeedForward when calling feed_forward_estimator.train(training_data):
ValueError: could not broadcast input array from shape (4) into shape (1)

An idea to make SimpleFeedForward work even when the ListDataset has unusable fields would be to maybe remove specific fields in the create_transformation method of SimpleFeedForwardEstimator using the RemoveFields class as in the DeepAREstimator class.
What do you think ? Would it make sense to implement that ?