Incomplete description of LSTM.__call__() params

[habernal]

URL(s) with the issue:

tensorflow/tensorflow/python/keras/layers/recurrent.py

Line 2461 in 48af54a

Call arguments:

Description of issue (what needs changing):

The current documentation reads:

  Call arguments:
    inputs: A 3D tensor.
    mask: Binary tensor of shape `(samples, timesteps)` indicating whether
      a given timestep should be masked.
    training: Python boolean indicating whether the layer should behave in
      training mode or in inference mode. This argument is passed to the cell
      when calling it. This is only relevant if `dropout` or
      `recurrent_dropout` is used.
    initial_state: List of initial state tensors to be passed to the first
      call of the cell.

It would be worth mentioning that

• mask, training, and initial_state are optional
• if initial_state is not provided, default zeros are imputed (by calling cell.get_initial_state() or cell.zero_state()), see tensorflow_core/python/ops/rnn.py lines 677 or 1382.

Due to iheritance of RNN layers (incl. mixins) and missing documentation at some places (haven't found any comments on get_initial_state), it's quite hard to figure it out.

Submit a pull request?

I'd like to keep it to someone who really knows the Keras internals; to me it's still a bit cryptic (i.e., where exactly get_initial_state_fn comes from, etc.)

[lamberta]

Thanks. That file is here, can you send a pull request?

[habernal]

@lamberta I'm not sure what you mean

[habernal]

Fixed in this PR: #31646