PhiFlow

Φ_Flow Installation

Requirements

• Python 3.6 or newer (e.g. Anaconda)
• pip (included in many Python distributions)

For GPU acceleration, deep learning and optimization, either PyTorch, TensorFlow 2.x or Jax must be registered with your Python installation. Note that these also require a CUDA installation with cuDNN libraries for GPU execution. We recommend CUDA 11.0 with cuDNN 8.

Installing Φ_Flow using pip

Note: If you want to use the Φ_Flow CUDA operations with TensorFlow, you have to build Φ_Flow from source instead (see below).

The following command installs the latest stable version of Φ_Flow with GUI support using pip.

$ pip install phiflow

To install the latest developer version of Φ_Flow, run

$ pip install --upgrade git+https://github.com/tum-pbs/PhiFlow@develop

Install PyTorch, TensorFlow or Jax in addition to Φ_Flow to enable machine learning capabilities and GPU execution.

To enable the web UI, also install dash.

$ pip install dash

If you only run Φ_Flow inside of Jupyter notebooks, dash is not required as Matplotlib will be used for plots and animations by default.

Installing Φ_Flow from Source

The Φ_Flow source additionally contains demo scripts and tests. In particular, it includes tests/verify.py, which tests your configuration.

Clone the git repository by running

$ git clone https://github.com/tum-pbs/PhiFlow.git <target directory>

This will create the folder <target directory> and copy all Φ_Flow source files into it.

With this done, you may compile CUDA kernels for better performance, see below.

Finally, Φ_Flow needs to be added to the Python path. This can be done in one of multiple ways:

• Marking <target directory> as a source directory in your Python IDE.
• Manually adding the cloned directory to the Python path.
• Installing Φ_Flow using pip: $ pip install <target directory>/. This command needs to be rerun after you make changes to the source code.

Compiling the CUDA Kernels for TensorFlow

The Φ_Flow source includes several custom CUDA kernels to speed up certain operations when using TensorFlow. To use these, you must have a TensorFlow compatible CUDA SDK with cuDNN as well as a compatible C++ compiler installed. We strongly recommend using Linux with GCC 7.3.1 (apt-get install gcc-4.8) for this. See the tested TensorFlow build configurations.

To compile the CUDA kernels for TensorFlow, clone the repository as described above, then run $ python <target directory>/setup.py tf_cuda. This will add the compiled CUDA binaries to the <target directory>.

Verifying the installation

If you have installed Φ_Flow from source, execute the included verify.py script.

$ python <Φ-Flow directory>/tests/verify.py

Otherwise, run the following Python code.

import phi; phi.verify()

Or from the command line:

$ python3 -c "import phi; phi.verify()"

If Φ_Flow and dependencies are installed correctly, you should see the Φ_Flow version, followed by additional information on the components at the end of the console output.

Unit tests

PyTest is required to run the unit tests. To install it, run $ pip install pytest.

Some unit tests require TensorFlow and PyTorch. Make sure, both are installed before running the tests.

Execute $ pytest <target directory>/tests/commit to run the normal tests. The result of this should match the automated tests run by Travis CI which are linked on the main GitHub page.

You can run the additional tests

$ pytest <target directory>/tests/gpu
$ pytest <target directory>/tests/release

This site is open source. Improve this page.