Unitary neural network is able to solve gradient vanishing and gradient explosion problem and help learning long term dependency. EUNN is an efficient and strictly enforced unitary parametrization based on SU(2) group. This repository contains the implementation of Efficient Unitary Neural Network(EUNN) in tensorflow.
If you find this work useful, please cite arXiv:1612.05231.
I am working on submitting this code to
tf.contrib so that in the future you can use it directly from official tensorflow.
requires TensorFlow > 1.2.0
To use EUNN in your model, simply copy eunn.py.
Then you can use EUNN in the same way you use built-in LSTM:
from eunn import EUNNCell cell = EUNNCell(hidden_size, capacity, fft, complex)
Integer. Only works for tunable style.
True, EUNN is set to FFT style. Default is
True, EUNN is set to complex domain. Default is
- For complex domain, the data type should be
- For real domain, the data type should be
Example tasks for EUNN
Copying memory task and pixel-permuted MNIST task for RNN in the paper are shown here. Due to copyright issue, we cannot release TIMIT task.
Copying Memory Task
python copying_task.py --model eunn --T 200 --fft
Pixel-Permuted MNIST Task
python mnist_task.py --model eunn --iter 20000 --hidden 512 --complex False