Backend not switching dynamically between tensorflow and torch

[jobs-git]

As the title says, using keras 3.8 on python 3.10:

from keras.src.utils import backend_utils
from keras.src import backend
dynamic = backend_utils.DynamicBackend ()
dynamic.set_backend('tensorflow')
print (backend.backend())
dynamic.set_backend('torch')
print (backend.backend())

tensorflow
tensorflow

[sonali-kumari1]

Hi @jobs-git -

Thanks for reporting this issue!
keras.backend.backend() returns global keras backend, which is tensorflow by default. However, DynamicBackend creates a local backend instance, so you can try printing dynamic.name to know the backend you are using locally with DynamicBackend like this:

from keras.src.utils import backend_utils
from keras.src import backend
import tensorflow as tf
dynamic = backend_utils.DynamicBackend()
dynamic.set_backend('tensorflow')
print(dynamic.name)
dynamic.set_backend('torch')
print(dynamic.name)

[jobs-git]

Right, its working that way. Could we get a documentation to on its usage? (i.e. Sequential among others) The backend_utils_test.py only shows one or few example of calculating with the backend using numpy.