Define custom LSTM with multiple inputs

Question

Following the tutorial writing custom layer, I am trying to implement a custom LSTM layer with multiple input tensors. I am providing two vectors input_1 and input_2 as a list [input_1, input_2] as suggested in the tutorial. The single input code is working but when I change the code for multiple inputs, its throwing the error,

self.kernel = self.add_weight(shape=(input_shape[0][-1], self.units),
    TypeError: 'NoneType' object is not subscriptable.

What change I have to do to get rid of the error? Here is the modified code.

class MinimalRNNCell(keras.layers.Layer):
    def __init__(self, units, **kwargs):
        self.units = units
        self.state_size = units
        super(MinimalRNNCell, self).__init__(**kwargs)
    def build(self, input_shape):
        print(type(input_shape))
        self.kernel = self.add_weight(shape=(input_shape[0][-1], self.units),
                                      initializer='uniform',
                                      name='kernel')
        self.recurrent_kernel = self.add_weight(
            shape=(self.units, self.units),
            initializer='uniform',
            name='recurrent_kernel')
        self.built = True
    def call(self, inputs, states):
        prev_output = states[0]
        h = K.dot(inputs[0], self.kernel)
        output = h + K.dot(prev_output, self.recurrent_kernel)
        return output, [output]   

# Let's use this cell in a RNN layer:
cell = MinimalRNNCell(32)
input_1 = keras.Input((None, 5))
input_2 = keras.Input((None, 5))
layer = RNN(cell)
y = layer([input_1, input_2])

Cynthea

Answer 1

Error is because of the line, y = layer([input_1, input_2]).

Replacing that line with y = layer((input_1, input_2)) (passing as Tuple of Inputs rather than List of Inputs), will resolve the error.

Complete working code using tf.keras is shown below:

from tensorflow import keras
from tensorflow.keras import backend as K
from tensorflow.keras.layers import RNN
import tensorflow as tf

class MinimalRNNCell(tf.keras.layers.Layer):
    def __init__(self, units, **kwargs):
        self.units = units
        self.state_size = units
        #self.state_size = [tf.TensorShape([units])]
        super(MinimalRNNCell, self).__init__(**kwargs)
    def build(self, input_shape):
        print(type(input_shape))
        self.kernel = self.add_weight(shape=(input_shape[0][-1], self.units),
                                      initializer='uniform',
                                      name='kernel')
        self.recurrent_kernel = self.add_weight(
            shape=(self.units, self.units),
            initializer='uniform',
            name='recurrent_kernel')
        self.built = True
    def call(self, inputs, states):
        prev_output = states[0]
        h = K.dot(inputs[0], self.kernel)
        output = h + K.dot(prev_output, self.recurrent_kernel)
        return output, [output]   

# Let's use this cell in a RNN layer:
cell = MinimalRNNCell(32)
input_1 = tf.keras.Input((None, 5))
input_2 = tf.keras.Input((None, 5))
layer = RNN(cell)
y = layer((input_1, input_2))

Output of the above code is:

<class 'tuple'>

Hope this helps. Happy Learning!

Hawthorne