TensorFlow Compression (TFC) contains data compression tools for TensorFlow.
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 (images, features, examples, etc.) while only sacrificing a small fraction of model performance. Take a look at the lossy data compression tutorial or the model compression tutorial to get started.
For a more in-depth introduction from a classical data compression perspective, consider our paper on nonlinear transform coding, or watch @jonycgn’s talk on learned image compression. For an introduction to lossy data compression from a machine learning perspective, take a look at @yiboyang’s review paper.
The library contains (see the API docs for details):
Range coding (a.k.a. arithmetic coding) implementations in the form of flexible TF ops written in C++. These include an optional “overflow” functionality that embeds an Elias gamma code into the range encoded bit sequence, making it possible to encode alphabets containing the entire set of signed integers rather than just a finite range.
Entropy model classes which simplify the process of designing rate–distortion optimized codes. During training, they act like likelihood models. Once training is completed, they encode floating point tensors into optimized bit sequences by automating the design of range coding tables and calling the range coder implementation behind the scenes.
Additional TensorFlow functions and Keras layers that are useful in the context of learned data compression, such as methods to numerically find quantiles of density functions, take expectations with respect to dithering noise, convolution layers with more flexible padding options and support for reparameterizing kernels and biases in the Fourier domain, and an implementation of generalized divisive normalization (GDN).
Important update: As of February 1, 2024, TensorFlow Compression is in maintenance mode. This means concretely:
The full feature set of TFC is frozen. No new features will be developed, but the repository will receive maintenance fixes.
Going forward, new TFC packages will only work with TensorFlow 2.14. This is due to an incompatibility introduced in the Keras version shipped with TF 2.15, which would require a rewrite of our layer and entropy model classes.
To ensure existing models can still be run with TF 2.15 and later, we are releasing a new package tensorflow-compression-ops, which only contains the C++ ops. These will be updated as long as possible for newer TF versions.
Binary packages are provided for both options on pypi.org: TFC and TFC ops.
Refer to the API documentation for a complete description of the classes and functions this package implements.
Please post all questions or comments on Discussions. Only file Issues for actual bugs or feature requests. On Discussions, you may get a faster answer, and you help other people find the question or answer more easily later.
Note: Precompiled packages are currently only provided for Linux and Darwin/Mac OS. To use these packages on Windows, consider installing TensorFlow using the instructions for WSL2 or using a TensorFlow Docker image, and then installing the Linux package.
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 2. For versions compatible with TensorFlow 1, see our previous releases.
To install TFC via pip, run the following command:
python -m pip install tensorflow-compression
To test that the installation works correctly, you can run the unit tests with:
python -m tensorflow_compression.all_tests
Once the command finishes, you should see a message OK (skipped=29) or
similar in the last line.
You can try out TFC live in a Colab. The following command installs the latest version of TFC that is compatible with the installed TensorFlow version. Run it in a cell before executing your Python code:
%pip install tensorflow-compression~=$(pip show tensorflow | perl -p -0777 -e 's/.*Version: (\d+\.\d+).*/\1.0/sg')
Note: The binary packages of TFC are tied to TF with the same minor version
(e.g., TFC 2.9.1 requires TF 2.9.x), and Colab sometimes lags behind a few days
in deploying the latest version of TensorFlow. As a result, using pip install
tensorflow-compression naively might attempt to upgrade TF, which can create
problems.
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:latest bash -c \
"python -m pip install tensorflow-compression &&
python -m tensorflow_compression.all_tests"
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. To
solve this, always install TensorFlow via pip rather than conda. For
example, this creates an Anaconda environment with CUDA libraries, and then
installs TensorFlow and TensorFlow Compression:
conda create --name ENV_NAME python cudatoolkit cudnn
conda activate ENV_NAME
python -m pip install tensorflow-compression
Depending on the requirements of the tensorflow pip package, you may need to
pin the CUDA libraries to specific versions. If you aren’t using a GPU, CUDA is
of course not necessary.
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 several implementations of published image compression models to enable
easy experimentation. Note that in order to reproduce published results, more
tuning of the code and training dataset may be necessary. Use the tfci.py
script above to access published models.
The following instructions talk about a re-implementation of the model published in:
“End-to-end optimized image compression”
J. Ballé, V. Laparra, E. P. Simoncelli
https://arxiv.org/abs/1611.01704
Note that the models directory is not contained in the pip package. The models
are meant to be downloaded individually. Download the file bls2017.py and run:
python bls2017.py -h
This will list the available command line options for the implementation. Training can be as simple as the following command:
python bls2017.py -V train
This will use the default settings. Note that unless a custom training dataset
is provided via --train_glob, the
CLIC dataset will be
downloaded using TensorFlow Datasets.
The most important training 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=/tmp/train_bls2017 &
When training has finished, the Python script saves the trained model to the
directory specified with --model_path (by default, bls2017 in the current
directory) in TensorFlow’s SavedModel format. The script can then be used to
compress and decompress images as follows. The same saved model 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).
To be compatible with the official TensorFlow pip package, the TFC pip package must be linked against a matching version of the C libraries. For this reason, to build the official Linux pip packages, we use these Docker images and use the same toolchain that TensorFlow uses.
Inside the Docker container, the following steps need to be taken:
tensorflow/compression repo from GitHub.tools/build_pip_pkg.sh inside the cloned repo.For example:
git clone https://github.com/tensorflow/compression.git /tensorflow_compression
docker run -i --rm \
-v /tmp/tensorflow_compression:/tmp/tensorflow_compression\
-v /tensorflow_compression:/tensorflow_compression \
-w /tensorflow_compression \
-e "BAZEL_OPT=--config=manylinux_2_17_x86_64" \
tensorflow/build:latest-python3.10 \
bash tools/build_pip_pkg.sh /tmp/tensorflow_compression <custom-version>
For Darwin, the Docker image and specifying the toolchain is not necessary. We just build the package like this (note that you may want to create a clean Python virtual environment to do this):
git clone https://github.com/tensorflow/compression.git /tensorflow_compression
cd /tensorflow_compression
BAZEL_OPT="--macos_minimum_os=10.14" bash \
tools/build_pip_pkg.sh \
/tmp/tensorflow_compression <custom-version>
In both cases, the wheel file is created inside /tmp/tensorflow_compression.
To test the created package, first install the resulting wheel file:
python -m pip install /tmp/tensorflow_compression/tensorflow_compression-*.whl
Then run the unit tests (Do not run the tests in the workspace directory where
the WORKSPACE file 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.all_tests
popd
When done, you can uninstall the pip package again:
python -m pip uninstall tensorflow-compression
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.
If you use this library for research purposes, please cite:
@software{tfc_github,
author = "Ballé, Johannes and Hwang, Sung Jin and Agustsson, Eirikur",
title = "{T}ensor{F}low {C}ompression: Learned Data Compression",
url = "http://github.com/tensorflow/compression",
version = "2.14.1",
year = "2024",
}
In the above BibTeX entry, names are top contributors sorted by number of commits. Please adjust version number and year according to the version that was actually used.
Note that this is not an officially supported Google product.