for multi layer LSTM model, how to specify certain layer to return_sequences as in keras?

[mw66]

Description

for multi layer LSTM model, how to specify certain layers to return_sequences as in keras?

https://keras.io/api/layers/recurrent_layers/lstm/

Use case

Currently the default NeuralForecast LSTM is fixed:

# Keras equivalent would be:
model = Sequential([
    LSTM(units, return_sequences=False),  # First layer - no sequences
    LSTM(units, return_sequences=True)   # Second layer - returns sequences
])

Is it possible to customize any individual layers to return_sequences=True, or units number to individually different numbers?

[marcopeix]

Currently, no, that is no possible. We can see if it's something the community wants.

[nasaul]

Adding to the comment of Marco.

If you need this functionality you could create a custom model for this one you will have to inherit from BaseModel and implement your own multi-layer LSTM with per-layer return control.