mmae 0.1.0

Package for multimodal autoencoders with Bregman divergences.

Package for multimodal autoencoders with Bregman divergences.

Description

This package contains an implementation of a flexible autoencoder that can take into account the noise distributions of multiple modalities. The autoencoder can be used to find a low-dimensional representation of multimodal data, taking advantage of the information that one modality provides about another.

Noise distributions are taken into account by means of Bregman divergences which correspond to particular exponential families such as Gaussian, Poisson or gamma distributions. Each modality can have its own Bregman divergence as loss function, thereby assuming a particular output noise distribution.

By default, the autoencoder network fusing multiple modalities consists of a variable number of ReLU layers that are densely connected. The number of layers and number of units per layer of the encoder and decoder networks are symmetric. Other network architectures can be easily implemented by overriding convenience methods.

Requirements

The package is compatible with Python 2.7 and 3.x and additionally requires NumPy, Six and Keras. It was tested with Python 2.7.15, Python 3.6.6, NumPy 1.15.2, Six 1.11.0 and Keras 2.2.4.

Installation

To install the mmae package, run:

pip install mmae

Usage

The main class of this package is MultimodalAutoencoder which is implemented in the module mmae.multimodal_autoencoder. This class can be used to easily construct a multimodal autoencoder for dimensionality reduction. The main arguments for instantiation are input_shapes which is a list of shapes for each modality, hidden_dims which is a list of the number of units per hidden layer of the encoder, output_activations which is a list of output activations for each modality, and losses which is a list of loss functions for each modality.

The last element of hidden_dims is the dimensionality of the latent space representation. The other elements are mirrored for the decoder construction. For instance, if hidden_dims = [128, 64, 8] then the encoder will have hidden layers with 128 and 64 units and output an 8 dimensional representation whereas the decoder will take the 8 dimensional representation, feed it into hidden layers with 64 and 128 units and produce multimodal outputs with shapes following input_shapes.

For losses, in addition to the standard Keras loss functions, regular Bregman divergences can be used. Current options are gaussian_divergence, gamma_divergence, bernoulli_divergence and poisson_divergence, corresponding to Gaussian, gamma, Bernoulli and Poisson noise models, respectively. To implement other divergences, additional classes can be derived from BregmanDivergence where the abstract methods _phi and _phi_gradient need to be overridden. BregmanDivergence is implemented in the mmae.bregman_divergences module.

The following code fits a multimodal autoencoder to MNIST, where the images are treated as one modality and the number label is treated as another modality:

from keras.datasets import mnist
from mmae.multimodal_autoencoder import MultimodalAutoencoder
# Load example data
(x_train, y_train), (x_test, y_test) = mnist.load_data()
# Scale pixel values to range [0, 1]
x_train = x_train.astype('float32') / 255.0
x_test = x_test.astype('float32') / 255.0
# Set network parameters
input_shapes = [x_train.shape[1:], (1,)]
# Number of units of each layer of encoder network
hidden_dims = [128, 64, 8]
# Output activation functions for each modality
output_activations = ['sigmoid', 'relu']
# Loss functions corresponding to a noise model for each modality
losses = ['bernoulli_divergence', 'poisson_divergence']
# Construct autoencoder network
autoencoder = MultimodalAutoencoder(input_shapes, hidden_dims,
                                    output_activations, losses)
# Train model where input and output are the same
autoencoder.fit([x_train, y_train], [x_train, y_train],
                epochs=100, batch_size=256,
                validation_data=([x_test, y_test], [x_test, y_test]))

To obtain a latent representation of the test data:

latent_test = autoencoder.encode([x_test, y_test])

To decode the latent representation:

reconstructed_test = autoencoder.decode(latent_test)

Encoding and decoding can also be merged into the following single statement:

reconstructed_test = autoencoder.predict([x_test, y_test])

The two modalities are fed directly into a dense fusion network. In order to preprocess each modality, for instance using a convolutional network for the image data, the MultimodalAutoencoder methods _construct_unimodal_encoders and _construct_unimodal_decoders can be overridden. These methods add networks between the input and the fusion encoder and between the fusion decoder and the output, respectively.

Source code

The source code of the mmae package is hosted on GitHub.

License

Copyright (C) 2018 Arno Onken

This file is part of the mmae package.

The mmae package is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3 of the License, or (at your option) any later version.

The mmae package is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with this program; if not, see <http://www.gnu.org/licenses/>.