This script is intended to assist in installing TensorFlow by source onto a Jetson TX# device
I highly recommend just using the pre-built wheel files. I build against Compute Capability 5.3 for TX1s, and 6.2 for TX2s. The result is a larger filesize compared to building against a single architecture, but the wheels are portable between TX1 and TX2.
From a terminal window navigate into the folder containing BuildTensorflow.sh
$ sudo ./BuildTensorflow.shThe build process should take between 4 and 6 hours depending on the performance of your device.
Note: Due to TensorFlow bug I am turning TF_NEED_TENSORRT=0 by default. If you are going to build against master you can open helperscript and set this value to 1.
Note: To build for Python 3, open helperscript and locate PYTHON_BIN_PATH=$(which python) and change this line to be PYTHON_BIN_PATH=$(which python3)
The BuildTensorflow.sh script will automatically install the Python wheel through pip install as well as place the C++ api objects into /usr/local/include and /usr/local/lib for you. However, if you wish to save the wheel file, it is located in /home/nvidia/TensorFlow_Install/tensorflow_pkg.
- Clean up folder permissions
- Somehow find a way to clean up swap if the build fails or is interrupted
- Include a command line option to build for Python 3
Licensed under the general MIT opensource license.
I do not warranty the use this software.
By using this script to build and install TensorFlow you are agreeing to the TensorFlow license, located
here.
The OpenJDK license is located here.
Bazel uses Apache License 2.0 and is located here.