How do you apply layer normalization in an RNN using tf.keras?

Question

I would like to apply layer normalization to a recurrent neural network using tf.keras. In TensorFlow 2.0, there is a LayerNormalization class in tf.layers.experimental, but it's unclear how to use it within a recurrent layer like LSTM, at each time step (as it was designed to be used). Should I create a custom cell, or is there a simpler way?

For example, applying dropout at each time step is as easy as setting the recurrent_dropout argument when creating an LSTM layer, but there is no recurrent_layer_normalization argument.

Answer 1

You can create a custom cell by inheriting from the SimpleRNNCell class, like this:

import numpy as np
from tensorflow.keras.models import Sequential
from tensorflow.keras.activations import get as get_activation
from tensorflow.keras.layers import SimpleRNNCell, RNN, Layer
from tensorflow.keras.layers.experimental import LayerNormalization

class SimpleRNNCellWithLayerNorm(SimpleRNNCell):
    def __init__(self, units, **kwargs):
        self.activation = get_activation(kwargs.get("activation", "tanh"))
        kwargs["activation"] = None
        super().__init__(units, **kwargs)
        self.layer_norm = LayerNormalization()
    def call(self, inputs, states):
        outputs, new_states = super().call(inputs, states)
        norm_out = self.activation(self.layer_norm(outputs))
        return norm_out, [norm_out]

This implementation runs a regular SimpleRNN cell for one step without any activation, then it applies layer norm to the resulting output, then it applies the activation. Then you can use it like that:

model = Sequential([
    RNN(SimpleRNNCellWithLayerNorm(20), return_sequences=True,
        input_shape=[None, 20]),
    RNN(SimpleRNNCellWithLayerNorm(5)),
])

model.compile(loss="mse", optimizer="sgd")
X_train = np.random.randn(100, 50, 20)
Y_train = np.random.randn(100, 5)
history = model.fit(X_train, Y_train, epochs=2)

For GRU and LSTM cells, people generally apply layer norm on the gates (after the linear combination of the inputs and states, and before the sigmoid activation), so it's a bit trickier to implement. Alternatively, you can probably get good results by just applying layer norm before applying activation and recurrent_activation, which would be easier to implement.