Unofficial implementation of ON-LSTM.
pip install keras-ordered-neuronsSame as LSTM except that an extra argument chunk_size should be given:
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Embedding, Bidirectional, Dense
from keras_ordered_neurons import ONLSTM
model = Sequential()
model.add(Embedding(input_shape=(None,), input_dim=10, output_dim=100))
model.add(Bidirectional(ONLSTM(units=50, chunk_size=5)))
model.add(Dense(units=2, activation='softmax'))
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy')
model.summary()Set recurrent_dropconnect to a non-zero value to enable drop-connect for recurrent weights:
from keras_ordered_neurons import ONLSTM
ONLSTM(units=50, chunk_size=5, recurrent_dropconnect=0.2)Set return_splits to True if you want to know the expected split points of master forget gate and master input gate.
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Embedding
from keras_ordered_neurons import ONLSTM
inputs = Input(shape=(None,))
embed = Embedding(input_dim=10, output_dim=100)(inputs)
outputs, splits = ONLSTM(units=50, chunk_size=5, return_sequences=True, return_splits=True)(embed)
model = Model(inputs=inputs, outputs=splits)
model.compile(optimizer='adam', loss='mse')
model.summary(line_length=120)
