TensorFlow has been fully updated to TF2. The official TensorFlow website shows that older versions of TensorFlow 1.15 only supported Python 3.7 and Cuda10, and that the official APT source for the new version of Ubuntu does not have Python 3.7 at all.

For newer versions of Python 3.8/ CUDA 11, if you want to use TF 1.15, one way to do this is to recompile or download someone else’s compiled binary WHL. But recompiling and finding the WHL is a bit of a problem. In order to enable new graphics card users to use TF1.15, Old Yellow cooperated with Google to maintain a library for TF1.15. The other way is to use Docker, but Docker is a little more cumbersome to use.

Library’s address is: https://github.com/NVIDIA/ten…

Usage:

# python3-m virtualenv venv # source venv/bin/activate PIP install --upgrade PIP install Nvidia-pyindex PIP install nvidia-tensorFlow [horovod] PIP install nvidia-tensorboard==1.15

Testing:

import tensorflow as tf
import tensorboard

tf.enable_eager_execution()
a = tf.random.uniform([1000, 1000])
b = tf.random.uniform([1000, 1000])
tf.matmul(a, b)

Just check the output is normal