Rules on setting num_feat_* for models

[Drewniacki]

What is the rule on setting num_feat_dynamic_real, num_feat_static_real and num_feat_static_cat (+ cardinality) when defining a model? Is it required to define it every time I want to create model with some external features? Or would model automatically read it from the dataset?

I am creating training and testing datasets via gluonts.dataset.common.ListDataset class:

train_ds = ListDataset(
    [
        {
            FieldName.TARGET: target,
            FieldName.START: start,
            FieldName.FEAT_DYNAMIC_REAL: fdr,
        }
        for (target, start, fdr) in zip(
            [ <target data here - shape: (n,m)> ],
            [ <window start here - shape: (n)> ],
            [ <dynamic feature here - shape (n,m,k)> ],
        )
    ],
    freq=freq,
)

Do I then need to also pass num_feat_dynamic_real = k when creating the model? What happens if I don't do that - does model disregard my dynamic features completely?

If I have to set num_feat_dynamic_real, num_feat_static_real and num_feat_static_cat, why is there no requirement to pass number of dynamic categorical features?

Thanks for your help!

[Drewniacki]

After some investigation I understood, that model uses gluonts.dataset.stat.calculate_dataset_statistics if num_feat_* are not passed to it during creation.

I have a question however. len(calculate_dataset_statistics(train_ds).feat_static_real == 23. But model parameter model.num_feat_static_real ==1 after I train it. And if I try to force it by setting DeepAREstimator(num_feat_static_real = 23) then the results come out super flat. Any idea what would be the reason for it?

Also another question still remains - why is there no requirement to pass number of dynamic categorical features?