State-Frequency Memory Recurrent Neural Network in TensorFlow and Keras
From Hu, Qi (2017), a state-frequency memory recurrent neural network implemented in TensorFlow and Keras. Implementation in Keras makes it easy to add additional recurrent layers of any type and experimentation with different loss functions and optimizers.
Adapted from the authors' Theano implementation.
Getting Started
Make sure you have the dependencies installed and the './datasets' folder in the same root directory. Can use the Keras and TensorFlow scripts independently or run the example scripts to see the network in action. The examples gather digits data from sklearn's digits data. You can also generate text sequence data for training by utilizing the other generation function in Data.py but due to the nature of the network, this will significantly increase computation time.
Also make sure you check out the original paper, very neat stuff.