NVlabs
Keras 3 compatibility
Hi, Thanks for your great contribution.
Sionna current version 0.18.0 is not implemented on Keras 3. However, keras 3 has many important features and it would be great to have sionna in keras 3 so that it can be integrated in models that are developed with Keras 3. Is this possible? Thanks
Hi @hafezmg48,
There reason why Keras 3 is not supported is that it does not work with complex-valued Tensors. Once this issue is resolved (keras-team/keras#19872), Sionna will be compatible with Keras 3.
However, note that Sionna will only support the TF backend.
Hi @hafezmg48,
There reason why Keras 3 is not supported is that it does not work with complex-valued Tensors. Once this issue is resolved (keras-team/keras#19872), Sionna will be compatible with Keras 3.
However, note that Sionna will only support the TF backend.
@jhoydis Thanks for considering this issue. Although we have discussed this and you correctly mentioned that the reason is that Keras 3 is not supporting complex tensors, I would like to also mention another related problem for your future reference. That is, the current supported TF version 2.15 is only compatible with at most the CUDA 11 which is currently outdated for many other AI projects which makes the integration of Sionna to those very hard (if we were to use the gpu for sionna). I believe this issue also has the same root as what you mentioned that TF =>2.16 works with keras 3 and keras 3 is not supporting complex-tensor, but wanted to specifically mention it for the reference. Thank you.
Closing as support for TensorFlow >2.15 has been resolved with the release of Sionna v1.0. Keras is no longer a required dependency.
