Incorrect Handling of out_features in RecurrentModel

[JesusPinedaC]

RecurrentModel inherits from RecurrentNeuralNetwork. In this model, the out_features argument is optional. However, in the __init__ method, the following call is made:

super().__init__(
    in_features,
    hidden_features=hidden_features[:-1],
    out_features=hidden_features[-1],
    batch_first=batch_first,
    return_cell_state=return_cell_state,
)

This implementation causes self.out_features to always be initialized with the last element of hidden_features, even when out_features is not explicitly provided by the user.

The problem

In the forward method, RecurrentModel checks for the existence of self.out_features. If present, the model attempts applies a dense layer (head).

if self.out_features is not None:
      outputs = self.head(outputs)
       ...
      return outputs

However, the presence of self.out_features is enforced, leading to a scenario where the model attempts to apply a dense head that was not intended by the user, as it is conditioned on a user-defined out_features that may not exist.

One simple solution might be to change the argument name to "model_out_features" but that would disrupt naming consistency in deepplay.

The following code snippet reproduces the issue:

rnn = dl.RecurrentModel(
    in_features=16, 
    hidden_features=[2],
    rnn_type="LSTM",
)
rnn = rnn.create()

x=torch.rand((32, 20, 16))
o = rnn(x)