Skip to contents

Set up Keras and TensorFlow

Source: vignettes/install-keras.qmd

The pixel-based workflow uses keras3 with a TensorFlow backend. Installing patternscaper also installs the R package keras3, but the Python environment and TensorFlow backend need a separate setup step.

This guide covers only what is needed to run the pixel-based workflow. For platform-specific options, including GPU support, see the official keras3 installation reference.

Choose Python 3.12

patternscaper currently recommends Python 3.12 for the pixel-based workflow. Other Python versions can remain installed, but the r-keras environment should use Python 3.12.

Start a fresh R session and check whether reticulate can find an existing Python 3.12 installation:

reticulate::virtualenv_starter(
  version = "3.12",
  all = TRUE
)

If this prints at least one Python installation, continue to Install Keras and TensorFlow. If it returns an empty result, install the latest Python 3.12 patch release:

reticulate::install_python("3.12:latest")

This installation is separate from other Python versions already on the computer. If Python 3.12 is installed in a non-standard location but is not found, see the official reticulate Python selection reference.

Install Keras and TensorFlow

In the same fresh R session, run:

keras3::install_keras(
  backend = "tensorflow",
  python_version = "3.12",
  restart_session = FALSE
)

reticulate::py_install(
  packages = "numpy<2",
  envname = "r-keras",
  pip = TRUE
)

The Keras installer creates a Python virtual environment named r-keras and installs Keras, TensorFlow, and the required Python dependencies. py_install installs a NumPy version compatible with the current keras3 requirements. The download and installation can take several minutes.

When the installation finished, restart R.

Verify the backend

In the fresh R session, load patternscaper and then ask Keras which backend it is using:

library(patternscaper)
keras3::config_backend()
#> [1] "tensorflow"
reticulate::py_config()

config_backend() should return "tensorflow". The Python configuration should point to the r-keras virtual environment, report Python 3.12 and a NumPy version below 2, and produce no warning about unsatisfied requirements declared through py_require().

TensorFlow may print informational oneDNN messages during this step which do not indicate installation failure.

Getting more help

If backend verification still fails, consult the official Keras installation guide and the keras3 installation reference for current setup instructions.

Python compatibility changes over time so consult the current TensorFlow installation guide before selecting a different Python version.

Once verification succeeds, continue with Classify landscapes with pixels.