Does tensorflow < 2.16 depend on keras or tf-keras?

[chenmoneygithub]

Hi team,

I have a question about Keras dependency of installing Tensorflow < 2.16. Basically if I run pip install tensorflow==2.15.0, it's going to install keras==2.15 as a dependency. Then when I upgrade to keras==3.0.2, it throws an error:

tensorflow-macos 2.15.0 requires keras<2.16,>=2.15.0, but you have keras 3.0.2 which is incompatible.

I was assuming older Tensorflow version will port to tf-keras instead of keras, but after I do pip install tf-keras, then from tensorflow import keras, the keras is still the multi-backend Keras. Could you help clarify the dependency? Will old tensorflow point to tf-keras?

For some context, I am trying to add keras 3 support to MLflow: mlflow/mlflow#10830, and the CI is failing because of incompatibility between Tensorflow version and Keras version. Thanks!

Best,
chenmoney

[sachinprasadhs]

tensorflow 2.15 depends on tf-keras/keras 2.15, but it is compatible with Keras 3 as well.

You can install tensorflow 2.15 first, which installs keras 2.15, you can update the keras package again to upgrade it to latest version, which is 3.0.2.

From tensorflow 2.16, it will install keras 3 by default.

Here is the commit for updating the requirement for tf-nightly(2.16)

[chenmoneygithub]

@sachinprasadhs Thanks!

tensorflow 2.15 depends on tf-keras/keras 2.15

What does it mean? tf-keras is a different package from keras, though they share the same version number. Now TF 2.15 is pointing to Keras instead of tf-keras. For more context, if I have both tf-keras==2.15 and keras==3.0.2 installed on my conda environment, then tensorflow==2.15.0 is using the keras==3.0.2 now. I am wondering if this is the expected behavior?

[sachinprasadhs]

tensorflow 2.15 by default installs keras 2.15 here https://github.com/keras-team/keras/tree/v2.15.0 and it will point to keras 2 .

If you install Keras 3 and if you also want the keras 2 behavior, then you can install tf-keras 2.15 and set the env variable with TF_USE_LEGACY_KERAS=1

More details on this can be found here https://keras.io/getting_started/#tensorflow--keras-2-backwards-compatibility

[chenmoneygithub]

@sachinprasadhs Thanks!

One more question, on keras.io it says:

Should you want tf.keras to stay on Keras 2 after upgrading to TensorFlow 2.16+, you can configure your TensorFlow installation so that tf.keras points to tf_keras. To achieve this:

Does this TF_USE_LEGACY_KERAS environment variable also work for older TF versions like tf 2.15?

[sachinprasadhs]

Ideally you don't need to use TF_USE_LEGACY_KERAS in 2.15, if you are using keras 2.15.
But, if you have installed keras 3 with tensorflow 2.15, use can use TF_USE_LEGACY_KERAS by installing tf-keras 2.15

[chenmoneygithub]

@sachinprasadhs Thanks! One more question - why should there be an error message tensorflow-macos 2.15.0 requires keras<2.16,>=2.15.0, but you have keras 3.0.2 which is incompatible. while Keras 3 works fine with TF 2.15? Keras 3 was developed and tested against TF 2.13 originally.

[sachinprasadhs]

I guess it is because of all the dependency in the same requirement.txt file.

If that is the case, try installing keras separately else you may find below error.

ERROR: Cannot install KERAS==3.0.2 and tensorflow==2.15.0 because these package versions have conflicting dependencies.

The conflict is caused by:
    The user requested KERAS==3.0.2
    tensorflow 2.15.0 depends on keras<2.16 and >=2.15.0

To fix this you could try to:
1. loosen the range of package versions you've specified
2. remove package versions to allow pip attempt to solve the dependency conflict

ERROR: ResolutionImpossible: for help visit https://pip.pypa.io/en/latest/topics/dependency-resolution/#dealing-with-dependency-conflicts

[github-actions]

This issue is stale because it has been open for 14 days with no activity. It will be closed if no further activity occurs. Thank you.

[github-actions]

This issue was closed because it has been inactive for 28 days. Please reopen if you'd like to work on this further.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No

[ShahriarAlom]

I have tried to ignore keras when installing Object detection and other Api's for version incompatibility and other issues, and after installation, I have seen that tasks include tensorflow and keras both, it doesn't works If keras wasn't installed properly with version control. And you can change keras version any time, where tf.keras will not help you in that case. So most of the installation and code will not work if Keras isn't installed independently.

[ShahriarAlom]

I will suggest to install keras 2.15.0 when you are installing any api or proto, after installation, install the latest any version you need, like 3.1.1 is latest. then older version will be auto replaced.