Content

Extends the standard keras LSTM and GRU layer with layer normalization, as described in here https://arxiv.org/pdf/1607.06450.pdf .

Example usage

The layers can be easily used like the normal layers:

from LayerNormalizationRNN import LSTM, GRU

inputs = Input(shape=(maxlen,))
x = Embedding(max_features, 128)(inputs)

x = LSTM(64, layer_to_normalize=("input", "output", "recurrent"), normalize_seperately=True)(x)
# x = GRU(64, layer_to_normalize=("input_gate", "input_recurrent", "recurrent_gate", "recurrent_recurrent"), normalize_seperately=True)(x)

predictions = Dense(1, activation='sigmoid')(x)


model = Model(inputs=inputs, outputs=predictions)
model.compile(optimizer='rmsprop',
              loss='binary_crossentropy',
              metrics=['accuracy'])

Explanation of arguments

The layer layer_to_normalize arguments specifies, after which matrix multiplication the layer normalization should be applied (see equations below).

The normalize_seperately argument specifies, whether the matrix multiplication for the forget, input, output... gates should be interpreted as one big one, or whether they should be split up in 4(LSTM)/2(GRU) smaller matrix multiplications, on which the layer normalization is applied.

Notes

For the LSTM layer, this implementation works with the implementaion=1 and implementation=0 flag. For GRU only implementation=1 is supported.

TODO

Test implementation on one of the experimental setups from the paper

Think of better names for the flags