keras-on-lstm 0.8.0

Unofficial implementation of ON-LSTM

Keras Ordered Neurons LSTM

[中文|English]

Unofficial implementation of ON-LSTM.

Install

pip install keras-ordered-neurons

Usage

Basic

Same as LSTM except that an extra argument chunk_size should be given:

from keras.models import Sequential
from 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()

DropConnect

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)

Expected Split Points

Set return_splits to True if you want to know the expected split points of master forget gate and master input gate.

from keras.models import Model
from 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)

tf.keras

Add TF_KERAS=1 to environment variables if you are using tensorflow.python.keras.