TF BB-ANS.
TensorFlow implementation of Bits Back with ANS by Townsend et al.
Presently, only the binarised MNIST version is provided. The implementation draws inspiration from the authors' code.
Both ELBO and compressed message length are ~0.17 bits per pixel (with perfect reconstruction).
To train the model:
$ python3 train.py -t
To evaluate the model:
$ python3 train.py
To run compression and decompression on e.g. 100 images:
$ python3 demo.py -n 100