Dynamic RNN In Keras: Use Custom RNN Cell To Track Other Outputs At Each Timestep

Is there a way to return multiple outputs for a given timestep when implementing a custom cell for an RNN in keras? E.g. outputs with shapes: (sequences=[batch, timesteps, hidden_u

Solution 1:

Figured it out. You can just make the output size a list with any dimensions and then the RNN will track the outputs. The class below also includes the use of constants in the RNN call because the previously mentioned paper passes an encoder latent space (z_enc) to the recurrent decoder:

class CustomMultiTimeStepGRUCell(tf.keras.layers.Layer):
    """Illustrates multiple sequence like (n, timestep, size) outputs."""

    def __init__(self, units, arbitrary_units, **kwargs):
        """Defines state for custom cell.
        
        :param units: <class 'int'> Hidden units for the RNN cell.
        :param arbitrary_units: <class 'int'> Hidden units for another
            dense network that outputs a tensor at each timestep in the
            unrolling of the RNN.
        """

        super().__init__(**kwargs)

        # Save args
        self.units = units
        self.arbitrary_units = arbitrary_units

        # Standard recurrent cell
        self.gru = tf.keras.layers.GRUCell(units=self.units)

        # For use with 'constant' kwarg in 'call' method
        self.concatenate = tf.keras.layers.Concatenate()
        self.dense_proj = tf.keras.layers.Dense(units=self.units)

        # For arbitrary computation at timestep t
        self.other_output = tf.keras.layers.Dense(units=self.arbitrary_units)

        # Hidden state size (i.e., h_t)...
        # it's useful to know in general that this refers to the following:
        # 'gru_cell = tf.keras.GRUCell(units=state_size)' 
        # 'seq, h_t = gru_cell(data)'
        # 'h_t.shape' -> '(?, state_size)'
        self.state_size = tf.TensorShape([self.units])

        # OUTPUT SIZE: PROBLEM SOLVED!!!!
        # This is the last dimension of the RNN sequence output.
        # Typically the last dimension matches the dimension of 
        # self.state_size, and in fact the keras RNN will infer 
        # the output size based on state size if output size is not
        # specified. In the case of output size that does not match the 
        # state size, you have to specify and in list format if 
        # multiple outputs can occur per timestep in the RNN.
        self.output_size = [tf.TensorShape([self.units]), tf.TensorShape([self.arbitrary_units])]

    def call(self, input_at_t, states_at_t, constants):
        """Forward pass for custom RNN cell.
        
        :param inputs_at_t: (batch, features) tensor from (batch, t, features)
            inputs
        :param states_at_t: <class 'tuple'> that has 1 element if
            if using GRUCell (h_t), or 2 elements if using LSTMCell (h_t, c_t)
        :param constants: <class 'tuple'> Unchanging tensors to be used
            in the unrolling of the RNN.

        :return: <class 'tuple'> with two elements.
            (1) <class 'list'> Both elements of this list are tensors
            that are tracked for each timestep in the unrolling of the RNN. 
            (2) Tensor representing the hidden state passed to the next
            cell.

            In the brief graphic below, a_t denotes the arbitrary output
            at each timestep. y_t = h_t_plus_1. x_t is some input at
            timestep t.

                    a_t  y_t
                     ^    ^
                   __|____|
            h_t    |      | h_t_plus_1
            -----> |      | ----------> .....
                   |______|
                      ^
                      |
                     x_t
         
            When all timesteps in x where x = {x_t}_{t=1}^{T} are processed
            by the RNN, the resulting shapes of the outputs assuming there 
            is only a single sample (batch = 1) would be the following:
            Y = (1, timesteps, units)
            A = (1, timesteps, arbitrary_units)
            h_t_plus_1 = (1, units)  # Last hidden state
            
            For a concrete example, see the end of this codeblock.
        """

        # Get correct inputs -- by default these args are tuples...
        # so you must index 0 to get the relevant element.
        # Note, if you are using LSTM, then the hidden states passed to the
        # the next cell in the RNN will be a tuple with two elements
        # i.e., (h_t, c_t) for the hidden and cell state, respectively.
        states_at_t = states_at_t[0]
        z_enc = constants[0]

        # Combine the states with z_enc
        combined = self.concatenate([states_at_t, z_enc])

        # Project to dimensions for GRU cell
        special_states_at_t = self.dense_proj(combined)

        # Standard GRU call
        output_at_t, states_at_t_plus_1 = self.gru(input_at_t, special_states_at_t)

        # Get another output at t
        arbitrary_output_at_t = self.other_output(input_at_t)

        # The outputs
        return [output_at_t, arbitrary_output_at_t], states_at_t_plus_1

# Dims
batch = 4
timesteps = 3
features = 12
latent = 8
hidden_units = 10
arbitary_units = 15

# Data
inputs = tf.random.normal(shape=(batch, timesteps, features))
h_t = tf.zeros(shape=(batch, hidden_units))
z_enc = tf.random.normal(shape=(batch, latent))

# An RNN cell to test multitimestep outputs
custom_multistep_cell = CustomMultiTimeStepGRUCell(units=hidden_units, arbitrary_units=arbitary_units)
custom_multistep_rnn = tf.keras.layers.RNN(custom_multistep_cell, return_sequences=True, return_state=True)

# Call cell
outputs, special_outputs, last_hidden = custom_multistep_rnn(inputs, initial_state=h_t, constants=z_enc)
print(outputs.shape)
print(special_outputs.shape)
print(last_hidden.shape)

>>> (4, 3, 10)
>>> (4, 3, 15)
>>> (4, 10)