Skip to content

ICSI - attempt at solving vanishing gradient without resorting to LSTMs

Notifications You must be signed in to change notification settings

rileyedmunds/Efficient-Unitary-RNNs

Repository files navigation

EUNN-tensorflow

Unitary neural network is able to solve gradient vanishing and gradient explosion problem and help learning long term correlation. Unitary RNN is promising to replace LSTM in multiple tasks. EUNN is an efficient unitary architecture based on SU(2) group. This repository contains an implementation of Efficient Unitary Neural Network(EUNN) and its Recurrent Neural Network implementation(EURNN).

If you find this work useful, please cite arXiv:1612.05231.

Installation

requires TensorFlow 1.2.0

Demo

./demo.sh

Usage

Use EUNN in RNN

To use EURNN in your model, simply copy EUNN.py and EURNN.py files.

Then you can use EURNN in the same way you use built-in LSTM:

from EURNN import EURNNCell
cell = EURNNCell(n_hidden, capacity=2, FFT=False, comp=False)

Args:

  • n_hidden: Integer. For FFT style, must be power of 2.
  • capacity: Optional. Integer. Only works for tunable style, must be even number.
  • FFT: Optional. Bool. If True, EURNN is set to FFT style. Default is False.
  • comp: Optional. Bool. If True, EURNN is set to complex domain. Default is False.

Note:

  • For complex domain, the data type should be tf.complex64
  • For real domain, the data type should be tf.float32

Use EUNN in other applications

To use EUNN in your model, simply copy EUNN.py file.

Then you can use EUNN in the following way:

from EUNN import EUNN
output = EUNN(input, capacity=2, FFT=False, comp=False)

Args:

  • input: 2D-Tensor. For FFT style, dimension must be power of 2.
  • capacity: Optional. Integer. Only works for tunable style, must be even number.
  • FFT: Optional. Bool. If True, EUNN is set to FFT style. Default is False.
  • comp: Optional. Bool. If True, EUNN is set to complex domain. Default is False.

Note:

  • For complex domain, the data type should be tf.complex64
  • For real domain, the data type should be tf.float32

Example tasks for EURNN

Two tasks for RNN in the paper are shown here. Use -h for more information

Copying Memory Task

requires: Model name (EURNN or LSTM);

optional parameters for the task:

delay time-T, number of iterations-I, batch size-B, hidden size-H;

optional parameters for EURNN:

capacity-L, complex or real-C, FFT style or tunable style-F.

Example:

python copying_task.py EURNN -T 100 -I 2000 -B 128 -H 128 -C True -F True

Pixel-Permuted MNIST Task

requires: Model name (EURNN or LSTM);

optional parameters for the task:

number of iterations-I, batch size-B, hidden size-H;

optional parameters for EURNN:

capacity-L, complex or real-C, FFT style or tunable style-F.

Example:

python mnist_task.py EURNN -I 2000 -B 128 -H 128 -L 4 -C True 

About

ICSI - attempt at solving vanishing gradient without resorting to LSTMs

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published