VPRNN

Official implementation and experimental results (excluding grid searches) for volume preserving recurrent neural networks (VPRNN) in keras with tensorflow 1.x backend.

Dependencies

This repository depends on the keras-vpnn package. Install that with

pip install git+http://github.com/wtaylor17/keras-vpnn

The remaining dependencies (of the python package) are:

1. wget
2. tensorflow 1.x, tested with 1.15.2
3. keras version 2.3.1
4. numpy

Additionally for some scripts, you may need matplotlib, seaborn, tqdm, or some other utilities. Some scripts import keras_layer_normalization (see here) but the experiments reported don't actually use it.

Installation

Run

pip install .

from the root directory of this repository, or alternatively

pip install git+http://github.com/wtaylor17/VPRNN

from anywhere with an internet connection.

Experiments

1. Addition problem T=500,1000,5000,10000
2. Sequential MNIST classification
3. Permuted MNIST classification
4. IMDB movie review classification
5. HAR-2 human activity classification

For each experiment (other than the addition problem), 5 training runs with the best hyperparameters were done after a grid search based on validation performance. The logs and models for the first of each of these 5 runs is provided.

Single layer accuracies and parameters (IMDB excludes embeddings):

Data Set	Test Accuracy (%)	Approx. Parameters
MNIST	98.12	11k
pMNIST	96.01	11k
IMDB	87.74	56k
HAR-2	94.94	5k

Known Issues

Attribute Error when Loading

If you get AttributeError: 'str' object has no attribute 'decode' when loading a VPRNN, try pip install h5py==2.10.0 --force-reinstall.

Loading IMDB Models

The pretrained model can't be easily loaded by default. Assuming embeddings are stashed, add the code

from vprnn.imdb_data import create_embeddings_matrix

class IMDBInit(keras.initializers.Initializer):
    def __init__(self, **kwargs):
        self.mat, *_ = create_embeddings_matrix()
        self.mat = K.constant(self.mat)
    
    def __call__(self, *args, **kwargs):
        return self.mat

setattr(keras.initializers, '<lambda>', IMDBInit)

before you call load_vprnn. An evaluation script may be added soon. For now, see this Colab notebook.

Strange Issue with HAR-2 on Linux

There was an issue with HAR data on linux that is now fixed. See this Colab notebook.