Closed tann9949 closed 3 years ago
Thanks a lot for this valuable contribution !
This actually prevents third-party backends such as plaidml for people who don't have the proper CUDA environment to run the tensorflow backend.
Before the change, I can simply set environment variable os.environ['KERAS_BACKEND'] = "plaidml.keras.backend"
to use plaidml instead of tensorflow with my AMD GPU
Keras is deprecated and using Keras package directly can cause a bug. If is safer to import keras as a tensorflow module