This is an openFrameworks addon for the TensorFlow 2 ML (Machine Learning) library. The code has been developed by the ZKM | Hertz-Lab as part of the project »The Intelligent Museum«.
Copyright (c) 2021-2023 ZKM | Karlsruhe.
BSD Simplified License.
For information on usage and redistribution, and for a DISCLAIMER OF ALL WARRANTIES, see the file, "LICENSE.txt," in this distribution.
Selected examples:
Style Transfer regular & arbitrary | Object Recognition | CharRnn (Frozen Graph) |
---|---|---|
Pose Estimation | Pix2Pix | Keyword Spotting |
---|---|---|
Video Matting |
---|
ofxTensorFlow2 is an openFrameworks addon for loading and running ML models trained with the TensorFlow 2 ML (Machine Learning) library:
TensorFlow is an end-to-end open source platform for machine learning. It has a comprehensive, flexible ecosystem of tools, libraries and community resources that lets researchers push the state-of-the-art in ML and developers easily build and deploy ML powered applications.
The addon utilizes the TensorFlow 2 C library wrapped by the open source cppflow C++ interface:
Run TensorFlow models in c++ without Bazel, without TensorFlow installation and without compiling TensorFlow. Perform tensor manipulation, use eager execution and run saved models directly from C++.
https://github.com/serizba/cppflow/
Additional classes wrap the process of loading & running a model and utility functions are provided for conversion between common openFrameworks types (images, pixels, audio samples, etc) and TensorFlow2 tensors.
openFrameworks is a cross platform open source toolkit for creative coding in C++.
Minimal quick start for a Unix shell to clone cppflow, download pre-built TensorFlow 2 dynamic libraries and pre-trained example models, starting in the root openFrameworks folder:
cd addons
git clone https://github.com/zkmkarlsruhe/ofxTensorFlow2.git
cd ofxTensorFlow2
git submodule update --init --recursive
./scripts/download_tensorflow.sh
./scripts/download_example_models.sh
For further information, please find detailed instructions below.
Note: The TensorFlow download script grabs the CPU-optimized build by default.
- openFrameworks
- Operating systems:
- Linux, 64-bit, x86_64
- macOS 10.14 (Mojave) or higher, 64-bit, x86_64 & arm64 (the latter via non-official builds)
- Windows, 64-bit, x86_64 (should work, not tested)
To use ofxTensorFlow2, first you need to download and install openFrameworks. The examples are developed against the latest release version of openFrameworks on http://openframeworks.cc/download.
Currently, ofxTensorFlow2 is being developed on Linux and macOS. Windows should work but has not yet been tested.
The main supported operating systems & architectures are those which have pre-built versions of libtensorflow available for download from the TensorFlow website. Other system configurations are possible but may require building and/or installing libtensorflow manually.
Clone (or download and extract) this repository to the addon folder of openFrameworks. Replace OF_ROOT with the path to your openFrameworks installation.
cd OF_ROOT/addons
git clone https://github.com/zkmkarlsruhe/ofxTensorFlow2.git
- TensorFlow 2
- cppflow 2
Since TensorFlow does not ship a C++ Library we make use of cppflow, which is a C++ wrapper around the TensorFlow 2 C API.
Pull cppflow to libs/cppflow
and checkout cppflow:
cd ofxTensorFlow2
git submodule update --init --recursive
Next, download the pre-built TensorFlow2 C library and extract the following folders to their destination. The lib files (.so, .dylib, .dll) must be placed within a subdirectory matching your system and/or build environment (osx, linux64, msys2, vs):
include/ --> libs/tensorflow/include
lib/ --> libs/tensorflow/lib/[osx/linux64/msys2/vs]
Note: If the libs are placed elsewhere, the OF ProjectGenerator will not find them and you will have linker errors when building. The naming is specified by openFrameworks.
To make this quick, you can use a script which automates the download:
./scripts/download_tensorflow.sh
As of summer 2024, the default download version will stay 2.8.0 as there are builds for all platforms. It is recommended to use a newer version of libtensorflow if is available or build from source. Check the TensorFlow2 C library page for the current release version.
By default, the script will try to auto-detect the system architecture. For example, on an Apple Silicon macOS system, the script will download builds for "arm64" while an Intel machine will use "x86_64".
When opting for GPU support set the TYPE
script variable:
TYPE=gpu ./scripts/download_tensorflow.sh
Additionally, to use a specific version, supply it as the first argument:
./scripts/download_tensorflow.sh 2.7.0
See https://www.tensorflow.org/install/gpu for more information on GPU support for TensorFlow.
As of Summer 2024, the TensorFlow website still does not provide arm64 builds for macOS. The pre-built versions downloaded by the script are provided by third parties on GitHub but they are not up to date and currently stop at version 2.8.0.
If you want the latest version of libtensorflow and you have some time, you can build it from source using the Makefile included in the libs
directory.
Instructions are found in libs/README.md
.
Note: After a successful build, make sure to run clean otherwise the OF Project Generator will look inside the libs/build
folder and add the dylibs twice. This will also take a long time, so best avoided.
To run applications using ofxTensorFlow2, the path to the addon's lib/tensorflow
subfolder needs to be added to the LD_LIBRARY_PATH
environment variable.
The path can be temporarily added via an export on the commandline (replace OF_ROOT
with the path to your openFrameworks installation) before running the application:
export LD_LIBRARY_PATH=OF_ROOT/addons/ofxTensorFlow2/libs/tensorflow/lib/linux64/:$LD_LIBRARY_PATH
make run
This step can also be automated by additional makefile targets provided by the addon_targets.mk
file. To use it, add the following to the end of the project's Makefile
:
# ofxTensorFlow2
include $(OF_ROOT)/addons/ofxTensorFlow2/addon_targets.mk
Additionally, this include can be added to an existing project by running the configure_makefile.sh
script with the path to the project directory as an argument:
scripts/configure_makefile.sh example_yolo_v4
This adds two additional targets, one for Debug and the other for Release, which run the application after exporting the LD_LIBRARY_PATH
. For example, to run a debug version of the application:
make RunDebugTF2
Similarly, for release builds use:
make RunReleaseTF2
For a permanent "set and forget" solution, the export line can be added to the end of your shell's user startup script, ie. ~/.zshrc
or /.bash_profile
to add the path whenever a new shell session is opened. Once set, the manual export is no longer required when running an ofxTensorFlow2 application.
To use libtensorflow installed to a system path, ie. by your system's package manager, the path(s) need to be added to the project header include and library search paths and the libraries need to be passed to the linker.
- If libtensorflow was downloaded to
libs/tensorflow/
, remove all files in this folder - Edit
addon_config.mk
under the "linux64" build target: comment the "local path" sections - If using the OF ProjectGenerator, (re)regenerate project files for projects using the addon
Note: When using libtensorflow installed to the system, the LD_LIBRARY_PATH
export is not needed.
The cppflow library requires C++14 (minimum) or C++17 which needs to be enabled when building on macOS.
As of openFrameworks 0.12, C++17 support is available by default.
libtensorflow is provided as pre-compiled dynamic libraries or can be built from source. On macOS these .dylib
files need to be configured and copied into the build macOS .app. These steps are automated via the scripts/macos_install_libs.sh
script and can be invoked when building, either by Xcode or the Makefiles.
Alternatively, you can use libtensorflow compiled and installed to the system, ie. /usr/local
or /usr/opt
. In this case, the dylibs do not need to be copied into the macOS .app, however the built app will not run on other computers without the same libraries installed to the same location.
After generating project files using the OF ProjectGenerator, add the following to one of the Run Script build phases in the Xcode project to invoke the macos_install_libs.sh
script, either via the configure_xcode.sh
script or manually.
Script method: close the project in Xcode if it's open, then run configure_xcode.sh
with the path to the project directory as argument:
scripts/configure_xcode.sh example_yolo_v4
Manual method:
- Select the project in the left-hand Xcode project tree
- Select the project build target under TARGETS
- Under the Build Phases tab, find the 2nd Run Script, and add the following
- OF 0.12: as the new last line (append), or
- OF 0.11: before the final
echo
line:
"$OF_PATH/addons/ofxTensorFlow2/scripts/macos_install_libs.sh" "$TARGET_BUILD_DIR/$PRODUCT_NAME.app";
Note: Whenever the project files are (re)generated, either method will need to be reapplied.
By default, Xcode Debug builds are for the current system architecture only. Build and run should work fine. Release builds, however, are "universal" and will build for all supported architectures, ie. "arm64" and "x86_64". This will cause linker errors since the libtensorflow builds used with this addon are generally single-architecture only.
To disable building for a specific architecture, it can be added to a list of "Excluded Architectures" within Xcode:
- Select the project in the left-hand Xcode project tree
- Select the project build target under TARGETS
- Under the Build Settings tab, find Exclude Architectures, double click on the second column, and add:
- on an Apple Silicon system: x86_64
- on an Intel system: arm64
When building an application using the makefiles, an additional step is required to install & configure the tensorflow2 dylibs into the project .app. This is partially automated by the scripts/macos_install_libs.sh
script which is called from the addon_targets.mk
file. To use it, add the following to the end of the project's Makefile
:
# ofxTensorFlow2
include $(OF_ROOT)/addons/ofxTensorFlow2/addon_targets.mk
Additionally, this include can be added to an existing project by running the configure_makefile.sh
script with the [ath to the project directory as an argument:
scripts/configure_makefile.sh example_yolo_v4
This adds two additional targets, one for Debug and the other for Release, which call the script to install the .dylibs. For example, to build a debug version of the application and install the libs, simply run:
make DebugTF2
Similarly, for release builds use:
make ReleaseTF2
This will also work when building the normal targets using two steps, for example:
make Debug
make DebugTF2
To use libtensorflow installed to a system path, ie. from a package manager like Homebrew, the path(s) need to be added to the project header include and library search paths and the libraries need to be passed to the linker. The scripts/macos_install_libs.sh
is not needed.
- If libtensorflow was downloaded to
libs/tensorflow/
, remove all files in this folder - Edit
addon_config.mk
under the "osx" build target:
- comment the "local path" sections and uncomment the "system path" sections
- If needed, change the path for your system, ie.
/usr/local
to/usr/opt
etc
- If using the OF ProjectGenerator, (re)regenerate project files for projects using the addon
In order to use the helper scripts, it is recommended to install the Msys2 distribution which provides both a Unix command shell and MinGW. Download the Msys2 "x86_64" 64 bit installer from: http://www.msys2.org/
In a Msys2 command shell, next install the curl
and unzip
commands:
pacman -S curl unzip
Now the scripts/download_tensorflow.sh
and scripts/example_models.sh
can be invoked.
The following steps were provided by Jonathan Frank Spring 2022 and were tested with Visual Studio 2022, CUDA 11.7, cuDNN 8.4.1.50, and libtensorflow 2.8.0. Help us expand this section as the main devs use Linux and macOS.
For best performance, it is suggested to install Nvidia CUDA for hardware acceleration.
- Download and install Nvidia CUDA (non-server) https://developer.nvidia.com/cuda-downloads
- Download and install Nvidia cuDNN (CUDA Deep Neural Network), requires dev program membership https://developer.nvidia.com/cudnn
- Add the following to
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v##.#\bin
- The contents of
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDNN\v#.#.#.#\bin
zlibwapi.dll
downloaded from http://www.winimage.com/zLibDll/zlib123dllx64.zip
- Install libtensorflow2 with GPU acceleration using either manually or via the
scripts/download_tensorflow.sh
Note: the cuDNN version seems to rely on specific CUDA and libtensorflow versions. Check the libtensorflow tested build configurations chart.
The example projects are located in the example_XXXX
directories.
Each example contains code to create a neural network and export it in the SavedModel format or previous frozen GraphDef format. Neural networks require training which may take hours or days in order to produce a satisfying output, therefore we provide pre-trained models which you can download as ZIP files, either from the release page on GitHub or from a public shared link here:
https://cloud.zkm.de/index.php/s/gfWEjyEr9X4gyY6
To make this quick, a script is provided to download and install the models for each example (requires a Unix shell, curl, and unzip):
cd OF_ROOT/addons/ofxTensorFlow2
./scripts/download_example_models.sh
By default, the example applications try to load a SavedModel named "model" (or "models" depending on the example) located in example_XXXX/bin/data/
. When downloading or training a model, please make sure the SavedModel is at this location and has the right name, otherwise update the model load path string.
Project files for the examples are not included so you will need to generate the project files for your operating system and development environment using the OF ProjectGenerator which is included with the openFrameworks distribution.
To (re)generate project files for an existing project:
- Click the "Import" button in the ProjectGenerator
- Navigate to the project's parent folder ie. "ofxTensorFlow2", select the base folder for the example project ie. "example_XXXX", and click the Open button
- Click the "Update" button
If everything went Ok, you should now be able to open the generated project and build/run the example.
Open the Xcode project, select the "example_XXXX Debug" scheme, and hit "Run".
For a Makefile build, build and run an example on the terminal:
cd example_XXXX
make ReleaseTF2
make RunRelease
For a Makefile build, build and run an example on the terminal:
cd example_XXXX
make Release
make RunReleaseTF2
Simply select ofxTensorFlow2 from the available addons in the OF ProjectGenerator before generating a new project. Make sure that all dependencies are installed and downloaded beforehand, otherwise the PG may miss some paths.
ofxTensorFlow2 works with the TensorFlow SavedModel format (preferred) and the older TensorFlow 1 frozen GraphDef format (legacy).
When referring to the "SavedModel" we mean the parent folder of the exported neural network containing two subfolder assets
and variables
and a saved_model.pb
file. Do not change anything inside this folder, however renaming the folder is permitted. Keep in mind to use the correct file path within the application.
Often you don't need or want to train your models from scratch. Therefor, you should take a look at the TF Hub. As TF2 is still rather new, there is not always a SavedModel for your purpose. Besides tfhub.dev you can search GitHub for a TF2 implementation of your model. A great place to start may be here. If you do not find a pre-trained model, it is still easier to run/extend the code of an existing project instead of starting from scratch.
If you happen to find a SavedModel that suits you, but actually don't know the in and output specifications, use the saved_model_cli
that comes with TensorFlow. For example:
saved_model_cli show --dir path/to/model/ --tag_set serve --signature_def serving_default
should give you the name and expected shape of the in and output tensors. If the names differ from the standard ones or you have more than one in or output tensor you can use ofxTF2Model::setup()
to specify them. This is also explained in the MultiIO example.
- Python 3
- Python Package Manger
- Virtual Environments (optional)
- Anaconda / Miniconda (includes all requirements)
- recent GPU (10+ series) + software support
- if no GPU is available try free services like Google"s Colab
We recommend using Python3 as Python2 is not being developed any longer. A python installation is usually extended using a package manager, e.g. pip or conda. To handle the dependencies of Python projects, virtual environments (venvs) are considered best practice. Most beginners to Python use Anaconda or the smaller Miniconda which have all of it to start with.
While you should not mix vens, you can do so for package managers e.g.:
conda install pip3
For each example project, create a new virtual environment. We will use conda
to do so:
cd example_XXXX/python
conda create -n myEnv python=3.7
conda activate myEnv
With our virtual environment set up and activated we need to install the required python packages. For each example we've listed the required packages using pip3 freeze > requirements.txt
. You can easily install them by running:
pip3 install -r requirements.txt
As the training procedure and the way of configuring it varies a lot between the examples, please refer to the README.md provided in the python
folder. Some may require to simply edit a config script and run:
python3 main.py
Some scipts may require to feed additional information to the main.py
script.
If you want to create your own Deep Learning project, here are some tips to guide you along the way.
Get an IDE (Integrated Development Environment) aka fancy text editor for development. As you will be using Python, choose a specialized IDE, e.g. Spyder (included in Anaconda) or PyCharm. Make sure to set the interpreter of the virtual environment for this project. If you chose to create the virtual environment using conda you will find a subfolder envs
in the installation folder of anaconda. This includes a folder for every virtual environment. Choose the right one and go to bin
and select the binary python as interpreter. This way the IDE can run and debug your projects.
Get familiar with Python. The official Python tutorial is a great place to start. Python has a lot of functions in its standard library, but there are a lot of other external packages to look out for:
- NumPy (efficient math algorithms and data structures)
- Matplotlib (plotting in the style of Matlab)
- TensorFlow 2 (ML library)
Get familiar with Keras. Since TensorFlow 2, Keras is the high level front-end of TensorFlow. It greatly reduces the effort of accessing common data structures (like labeled pictures), defining a neural network architecture and observing the training process using callbacks. Besides that, you can always call TensorFlow's core functions such as data pipelines.
Get some structure for your project. Your project could look a little bit like this:
data
: stores scripts to download and maybe process some datasrc
: contains Python code for the model, pre-processing and train, test and eval proceduresmain.py
: often serves as a front to call the train, eval or test scriptsconfig.py
: stores high level parameters such as learning rate, batch size, etc. Edit this file for different experiments. Formats other than .py are fine too, but it's very easy to integrate. It's a good choice to save this file along with trained models.requirements.txt
: contains required packages
Get familiar with Machine Learning concepts. There is plenty of free information out there! Here is a list of material to look into:
- Coursera: founded by ML expert Andrew Ng, lists free online courses for a lot of fields (including Python and Machine Learning)
- Deeplearning.ai: a website dedicated to Deep Learning - also founded by Andrew Ng
- Deep Learning book: a free website accompanying the book "Deep Learning" by Ian Goodfellow (known for GANs)
- Stanford CS231: YouTube playlist of Stanford's Computer Vision course CS231
- Machine Learning Mastery: a popular blog about practical ML techniques. It focuses on the ease of use
Get familiar with TensorFlow's Tutorials. Besides learning how to write TensorFlow code, the tutorials will teach you ML concepts like over- and underfitting. Another great place to start is this repository. It is a vast conglomeration of material related to TensorFlow 2.X.
Get to know common datasets. A great place to start is Kaggle. Here you can find thousands of datasets and accompanying code (in form of Python notebooks that run in your browser). TF datasets is also very popular as most datasets do not require manual download.
Get inspired and take the risk of making errors! We can not help you with the latter but check out this repo for some inspiration.
You can help develop ofxTensorFlow2 on GitHub: https://github.com/zkmkalrsruhe/ofxTensorFlow2
Create an account, clone or fork the repo, then request a push/merge.
If you find any bugs or suggestions please log them to GitHub as well.
Xcode will try to codesign the libtensorflow dylibs when building the app bundle and this step may fail if there is no code signing identity set, ie. "-" anonymous. The error will look something like: Command CodeSign failed with a nonzero exit code
.
If you haven't set a signing identity but want to run the app, disable this step in Build Phases -> Copy Files step:
The default OF-generated Xcode project will try to build for both arm64 and x86_64 in Release, so you may see linker errors such as:
Ignoring file ../../../addons/ofxTensorFlow2/libs/tensorflow/lib/osx/libtensorflow.2.4.0.dylib, building for macOS-arm64 but attempting to link with file built for macOS-x86_64
...
"_TFE_ContextOptionsSetConfig", referenced from:
To make the build succeed, you can exclude the arm64 architecture in Xcode so the project. See the info in the the "macOS / Xcode build" subsection under the "Installation and Build" section.
On macOS, the libtensorflow dynamic libraries (dylibs) need to be copied into the .app bundle. This error indicates the library loader cannot find the dylibs when the app starts and the build process is missing a step. Please check the "macOS" subsection under the "Installation and Build" section.
The pre-built libtensorflow downloaded to libs/tensorflow
comes with AVX (Advanced Vector Extensions) enabled which is an extension to the Intel x86 instruction set for fast vector math. CPUs older than circa 2013 may not support this and the application will simply crash with error such as:
in libtensorflow_framework.2.dylib
...
EXC_BAD_INSTRUCTION (code=EXC_I386_INVOP, subcode=0x0)
This problem may also be seen when using libtensorflow installed via Homebrew.
The only solution is to build libtensorflow from source with AVX disabled use a machine with a newer CPU. To check if your CPU supports AVX use:
# print all CPU features
sysctl -a | grep cpu.feat
# prints only if CPU supports AVX
sysctl -a | grep cpu.feat | grep AVX
Systems confirmed: Mac Pro (Mid 2012)
The pre-built libtensorflow dynamic libraries downloaded from the TensorFlow website require a minimum of macOS 10.14. On macOS 10.13 or lower, the project may build but will fail on run with a runtime loader error:
dyld: lazy symbol binding failed: Symbol not found: ____chkstk_darwin
Referenced from: /Users/na/of_v0.11.0_osx_release/addons/ofxTensorFlow2/example_basics/bin/example_basics.app/Contents/MacOS/./../Frameworks/libtensorflow.2.dylib (which was built for Mac OS X 10.15)
Expected in: /usr/lib/libSystem.B.dylib
The only solutions are:
- upgrade to macOS 10.14 or newer (easier)
- use libtensorflow compiled for your system:
- installed to system via a package manager, ie. Homebrew or Macports (harder)
- or, build libtensorflow manually (probably hardest)
An artistic-curatorial field of experimentation for deep learning and visitor participation
The ZKM | Center for Art and Media and the Deutsches Museum Nuremberg cooperate with the goal of implementing an AI-supported exhibition. Together with researchers and international artists, new AI-based works of art will be realized during the next four years (2020-2023). They will be embedded in the AI-supported exhibition in both houses. The Project „The Intelligent Museum” is funded by the Digital Culture Programme of the Kulturstiftung des Bundes (German Federal Cultural Foundation) and funded by the Beauftragte der Bundesregierung für Kultur und Medien (Federal Government Commissioner for Culture and the Media).
As part of the project, digital curating will be critically examined using various approaches of digital art. Experimenting with new digital aesthetics and forms of expression enables new museum experiences and thus new ways of museum communication and visitor participation. The museum is transformed to a place of experience and critical exchange.