This project contains data compression ops and layers for TensorFlow. The project website is at tensorflow.github.io/compression.

You can use this library to build your own ML models with end-to-end optimized data compression built in. It's useful to find storage-efficient representations of your data (features, examples, images, etc.) while only sacrificing a tiny fraction of model performance. It can compress any floating point tensor to a much smaller sequence of bits.

Specifically, the EntropyBottleneck class in this library simplifies the process of designing rate–distortion optimized codes. During training, it acts like a likelihood model. Once training is completed, it encodes floating point tensors into optimal bit sequences by automating the design of probability tables and calling a range coder implementation behind the scenes.

For an introduction to lossy image compression with machine learning, take a look at @jonycgn's talk on Learned Image Compression.

For usage questions and discussions, please head over to our Google group.

Refer to the API documentation for a complete description of the Keras layers and TensorFlow ops this package implements.

There's also an introduction to our EntropyBottleneck class here, and a description of the range coding operators here.

Note: Precompiled packages are currently only provided for Linux (Python 2.7, 3.3-3.6) and Darwin/Mac OS (Python 2.7, 3.7). To use these packages on Windows, consider using a TensorFlow Docker image and installing tensorflow-compression using pip inside the Docker container.

Set up an environment in which you can install precompiled binary Python packages using the pip command. Refer to the TensorFlow installation instructions for more information on how to set up such a Python environment.

The current version of tensorflow-compression requires TensorFlow 1.15. For TensorFlow 1.14 or earlier, see our previous releases. You can install TensorFlow from any source. To install it via pip, run the following command:

pip install tensorflow-gpu==1.15

for GPU support, or

pip install tensorflow==1.15

for CPU-only.

Then, run the following command to install the tensorflow-compression pip package:

pip install tensorflow-compression

To test that the installation works correctly, you can run the unit tests with

python -m tensorflow_compression.python.all_test

Once the command finishes, you should see a message OK (skipped=12) or similar in the last line.

To use a Docker container (e.g. on Windows), be sure to install Docker (e.g., Docker Desktop), use a TensorFlow Docker image, and then run the pip install command inside the Docker container, not on the host. For instance, you can use a command line like this:

docker run tensorflow/tensorflow:1.15.0-py3 bash -c \
    "pip install tensorflow-compression &&
     python -m tensorflow_compression.python.all_test"

This will fetch the TensorFlow Docker image if it's not already cached, install the pip package and then run the unit tests to confirm that it works.

It seems that Anaconda ships its own binary version of TensorFlow which is incompatible with our pip package. It also installs Python 3.7 by default, which we currently don't support on Linux. To solve this, make sure to use Python 3.6 on Linux, and always install TensorFlow via pip rather than conda. For example, this creates an Anaconda environment with Python 3.6 and CUDA libraries, and then installs TensorFlow and tensorflow-compression with GPU support:

conda create --name ENV_NAME python=3.6 cudatoolkit=10.0 cudnn
conda activate ENV_NAME
pip install tensorflow-gpu==1.15 tensorflow-compression

We recommend importing the library from your Python code as follows:

import tensorflow as tf
import tensorflow_compression as tfc

In the models directory, you'll find a python script tfci.py. Download the file and run:

python tfci.py -h

This will give you a list of options. Briefly, the command

python tfci.py compress <model> <PNG file>

will compress an image using a pre-trained model and write a file ending in .tfci. Execute python tfci.py models to give you a list of supported pre-trained models. The command

python tfci.py decompress <TFCI file>

will decompress a TFCI file and write a PNG file. By default, an output file will be named like the input file, only with the appropriate file extension appended (any existing extensions will not be removed).

The models directory contains an implementation of the image compression model described in:

"End-to-end optimized image compression"
J. Ballé, V. Laparra, E. P. Simoncelli
https://arxiv.org/abs/1611.01704

To see a list of options, download the file bls2017.py and run:

python bls2017.py -h

To train the model, you need to supply it with a dataset of RGB training images. They should be provided in PNG format. Training can be as simple as the following command:

python bls2017.py --verbose train --train_glob="images/*.png"

This will use the default settings. The most important parameter is --lambda, which controls the trade-off between bitrate and distortion that the model will be optimized for. The number of channels per layer is important, too: models tuned for higher bitrates (or, equivalently, lower distortion) tend to require transforms with a greater approximation capacity (i.e. more channels), so to optimize performance, you want to make sure that the number of channels is large enough (or larger). This is described in more detail in:

"Efficient nonlinear transforms for lossy image compression"
J. Ballé
https://arxiv.org/abs/1802.00847

If you wish, you can monitor progress with Tensorboard. To do this, create a Tensorboard instance in the background before starting the training, then point your web browser to port 6006 on your machine:

tensorboard --logdir=. &

When training has finished, the Python script can be used to compress and decompress images as follows. The same model checkpoint must be accessible to both commands.

python bls2017.py [options] compress original.png compressed.tfci
python bls2017.py [options] decompress compressed.tfci reconstruction.png

This section describes the necessary steps to build your own pip packages of tensorflow-compression. This may be necessary to install it on platforms for which we don't provide precompiled binaries (currently only Linux and Darwin).

We use the Docker image tensorflow/tensorflow:custom-op-ubuntu16 for building pip packages for Linux. Note that this is different from tensorflow/tensorflow:devel. To be compatible with the TensorFlow pip package, the GCC version must match, but tensorflow/tensorflow:devel has a different GCC version installed.

Inside a Docker container from the image, the following steps need to be taken.

1. Install the TensorFlow pip package.
2. Clone the tensorflow/compression repo from GitHub.
3. Run :build_pip_pkg inside the cloned repo.

For example:

sudo docker run -v /tmp/tensorflow_compression:/tmp/tensorflow_compression \
    tensorflow/tensorflow:custom-op-ubuntu16 bash -c \
    "pip install tensorflow==1.15 &&
     git clone https://github.com/tensorflow/compression.git
         /tensorflow_compression &&
     cd /tensorflow_compression &&
     bazel run -c opt --copt=-mavx :build_pip_pkg"

The wheel file is created inside /tmp/tensorflow_compression. Optimization flags can be passed via --copt to the bazel run command above.

To test the created package, first install the resulting wheel file:

pip install /tmp/tensorflow_compression/tensorflow_compression-*.whl

Then run the unit tests (Do not run the tests in the workspace directory where WORKSPACE of tensorflow_compression repo lives. In that case, the Python interpreter would attempt to import tensorflow_compression packages from the source tree rather than from the installed package system directory):

pushd /tmp
python -m tensorflow_compression.python.all_test
popd

When done, you can uninstall the pip package again:

pip uninstall tensorflow-compression

To build packages for Darwin (and potentially other platforms), you can follow the same steps, but the Docker image should not be necessary.

We provide evaluation results for several image compression methods in terms of different metrics in different colorspaces. Please see the results subdirectory for more information.

• Johannes Ballé (github: jonycgn)
• Sung Jin Hwang (github: ssjhv)
• Nick Johnston (github: nmjohn)
• David Minnen (github: minnend)

Note that this is not an officially supported Google product.