MNIST AutoEncoder

Output of the example script. First two columns show training set, second two columns show testing set.

This example shows how to train a simple auto-encoder for MNIST, with a quantized bottleneck, and a conditional probability model that estimates the distribution of the bottleneck. It basically follows recent image compression papers in terms of structure, but is much simpler and not tuned.

Goals:

• Show how to use the probability distribution predicted from some probability model CNN to estimate the bitrate
• Actually encode symbols with that model using torchac, and see how many bits it takes to store

What could be improved:

• The models don't actually train very well. The purpose of the example is to show how torchac can be used.

Runnig it

The example supports interactive plots with matplotlib, which it tries to use by default. If they are not available (in headless environments), it falls back to storing plots in plots. If you want to force non-interactive, you can set export NO_INTERACTIVE=1 before running the script.

# If not installed already:
pip install matplotlib
pip install torchac

# Got to this folder
cd examples/mnist_qutoencder

# Run the script
python mnist_autoencoder_example.py