- *Proper release schedule and binaries.* We've tagged existing code versions 1.0 and 1.1, which work without installing binaries with TensorFlow <=1.12 and 1.13, respectively. These versions rely on a binary range coder implementation in tf.contrib, which will not exist any more in TensorFlow 1.14 (the TensorFlow team will remove tf.contrib entirely in release 2.0). For this reason, tensorflow-compression 1.2 ships with its own range coder implementations. Pre-compiled pip packages will be provided for Linux initially. We are working with the TensorFlow team to also provide Darwin (Mac) binaries.
- *Support for conditional entropy models.* This version includes three new classes, `GaussianConditional`,`LaplacianConditional`, and `LogisticConditional`, which implement conditional entropy models (entropy models whose probabilities are computed using another neural network). These are necessary for building the hyperprior models published in [our ICLR 2018 paper](https://openreview.net/forum?id=rkcQFMZRb). This also includes a more flexible range coding implementation, which can encode unbounded integer values.
- *Support for trained models.* The library now includes support for a new file format, TFCI, and packaged metagraphs of models we've published. `examples/tfci.py` implements an easy-to-use command line interface for converting PNG images to TFCI and back.
- *More flexible SignalConv layers*. Among some other minor improvements, the `SignalConv*` Keras layers now implement rational up-/downsampling factors.
Upcoming features not yet implemented in this beta:
- More pre-trained models (equivalent to publications [End-to-end optimized image compression](https://openreview.net/forum?id=rJxdQ3jeg) as well as [Joint Autoregressive and Hierarchical Priors for Learned Image Compression](http://papers.nips.cc/paper/8275-joint-autoregressive-and-hierarchical-priors-for-learned-image-compression)).
- Example file for training your own model with a hierarchical prior (as in the ICLR 2018 paper).
- Native Darwin (Mac) binaries. We can't promise we'll get support ready for 1.2, but we will try.
- More (and better) documentation.