This repository contains a Keras implementation of the Independently Recurrent Neural Network (IndRNN, Li et al. [1]).

The file indrnn.py contains the classes:

• IndRNNCell, which construct the basic IndRNN model by slightly modifying Keras' SimpleRNNCell. Its recurrent kernel is a single-row matrix, which is multiplied element-wise at each timesteps with the model's previous output;
• IndRNN, which is a RNN with a single IndRNNCell;
• CuDNNIndRNN, which uses the CuDNN-RNN as a black-box to compute the IndRNN step by trasforming the recurrent kernel in a diagonal matrix. Warning: This is just a stub and it is recommended to use the IndRNN class instead.

The IndRNN uses by default a ReLU activation, its recurrent kernel is constrained and initialized with random values in the range (-1, 1). It can be used "as-is", as following

from keras.models import Sequential
from keras.layers import Dense
from indrnn import IndRNN

model = Sequential()
model.add(IndRNN(128, input_shape=(None, 2)))
model.add(Dense(1))

model.compile(optimizer='adam', loss='mse')
model.summary()

Layer (type)                 Output Shape              Param #   
=================================================================
ind_rnn_1 (IndRNN)           (None, 128)               512       
_________________________________________________________________
dense_1 (Dense)              (None, 1)                 129       
=================================================================
Total params: 641
Trainable params: 641
Non-trainable params: 0

If you want to control better the recurrent kernel values, as in [1], you can use Keras' initializers and contraints, as following

from keras.initializers import RandomUniform
from keras.constraints import MaxNorm

T = 5000
limit = 2**(1/T)

model = Sequential()
model.add(IndRNN(128, 
                 recurrent_initializer=RandomUniform(-limit, limit),
                 recurrent_constraint=MaxNorm(limit),
                 input_shape=(None, 2)))
model.add(Dense(1))

Keep in mind that Keras' constraints act by default on kernel's first axis. Since IndRNN has a 2-dimensional single-row recurrent kernel, applying MaxNorm constraint is the same as constraining the kernel's absolute values.

You can see the IndRNN in action in this Jupyter notebook.

[1] Li, S., Li, W., Cook, C., Zhu, C. and Gao, Y., 2018. Independently recurrent neural network (indrnn): Building a longer and deeper rnn. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 5457-5466).