Tensorflow/Keras code replicating the experiments in: https://arxiv.org/abs/1904.07199

@article{brekelmans2019exact,
  title={Exact Rate-Distortion in Autoencoders via Echo Noise},
  author={Brekelmans, Rob and Moyer, Daniel and Galstyan, Aram and Ver Steeg, Greg},
  journal={arXiv preprint arXiv:1904.07199},
  year={2019}
}

Echo noise is flexible, data-driven alternative to Gaussian noise that admits an simple, exact expression for mutual information by construction. Applied in the autoencoder setting, we show that regularizing with I(X:Z) corresponds to the optimal choice of prior in the Evidence Lower Bound and leads to significant improvements over VAEs.

For easy inclusion in other projects, the echo noise functions are included in one all-in-one file, echo_noise.py, which can be copied to a project and included directly, e.g.:

import echo_noise

There are two basic functions implemented, the noise function itself (echo_sample) and the MI calculation (echo_loss), both of which are included in echo_noise.py. Except for libaries, echo_noise.py has no other file dependencies.

Echo noise is meant to be used similarly to the Gaussian noise in VAEs, and was implemented with VAE implementations in mind. Assuming the inference network provides z_mean and z_log_scale, a Gaussian Encoder would look something like:

z = z_mean + tf.exp(z_log_scale) * tf.random.normal( tf.shape(z_mean) )

The Echo noise equivalent implemented here is:

z = echo_noise.echo_sample( [z_mean, z_log_scale] )

Similarly, VAEs often calculate a KL divergence penalty based on z_mean and z_log_scale. The Echo noise penalty, which is the mutual information I(x,z), can be computed using:

loss = ... + echo_noise.echo_loss([z_log_scale])

A Keras version of this might look like the following:

z_mean = Dense(latent_dim, activation = model_utils.activations.tanh64)(h)
z_log_scale = Dense(latent_dim, activation = tf.math.log_sigmoid)(h)
z_activation = Lambda(echo_noise.echo_sample)([z_mean, z_log_scale])
echo_loss = Lambda(echo_noise.echo_loss)([z_log_scale])

These functions are also found in the experiments code, model_utils/layers.py and model_utils/losses.py.

python run.py --config 'echo.json' --beta 1.0 --filename 'echo_example' --dataset 'binary_mnist'

Experiments are specifed using the config files, which specify the network architecture and loss functions. run.py calls model.py to parse these configs/ and create / train a model. You can also modify the tradeoff parameter beta, which is multiplied by the rate term, or specify the dataset using 'binary_mnist', 'omniglot', or 'fmnist'. . Analysis tools are mostly omitted for now, but the model loss training history is saved in a pickle file.

We can choose to sample training examples with or without replacement from within the batch for constructing Echo noise.
For sampling without replacement, we have two helper functions which shuffle index orderings for x^(l). permute_neighbor_indices sets the output batch_size != None and is much faster. indices_without_replacement maintains batch_size = None (e.g. for variable batch size or fitting with keras fit). Control these with set_batch option.

Also be wary of leftover batches : we choose d_max samples to construct Echo noise from within the batch, so small batches (especially without replacement) may give inaccurate noise.

We compare diagonal Gaussian noise encoders ('VAE') and IAF encoders, alongside several marginal approximations : standard Gaussian prior, standard Gaussian with MMD penalty (info_vae.json or iaf_prior_mmd.json), Masked Autoregressive Flow (MAF), and VampPrior. All combinations can be found in the configs/ folder.