Build Tensorflow in Automatic Mixed Precision Container

AI & Data Science Deep Learning (Training & Inference) Frameworks
1

Hi all!

Is it possible to rebuild Tensoflow from sources in Automatic Mixed Precision Containers (eg: nvcr.io/nvidia/tensorflow:19.06-py2) to speed up CPU operations and keep the FP16 compatibility?

Thanks!

2

The sources used to build the provided TF binaries are provided under /opt/tensorflow in the nvcr containers. You can tweak the sources as desired and then run /opt/tensorflow/nvbuild.sh to compile and install a modified version of TensorFlow.