Original source cloned from simonnanty's gist: https://gist.github.com/simonnanty/bf56dd6ae3cc9566cd3f004431ff512d

This repository contains an implementation of Minimal RNN by M. Chen, 2017

MinimalRNN: Toward More Interpretable and Trainable Recurrent Neural Networks https://arxiv.org/abs/1711.06788v1

Just import minimalrnn.py:

from minimalrnn import MinimalRNNCell

and then use MinimalRNNCell like any other RNNCell objects.

example-mnist-minimal-rnn.ipynb contains a simple MNIST RNN example, tried to show how to create a custom phi transformation function for the minimal-RNN.

Caveats: I couldn't use tf.layers.batch_normalization() in the custom phi transform because of following error:

InvalidArgumentError: The node 'gradients/rnn/TensorArrayStack/TensorArrayGatherV3_grad/TensorArrayGrad/TensorArrayGradV3' has inputs from different frames. The input 'rnn/while/rnn/minimal_rnn_cell/batch_normalization/AssignMovingAvg' is in frame 'rnn/while/while_context'. The input 'rnn/TensorArray' is in frame ''.

Any suggestion for above issue will be welcomed. :-)

Thank you for reading this.