similar functions for from_tensor to_tensor from ragged api

[innat]

I think ragged doesn't support yet. But is there any way to handle such following cases?

tf.RaggedTensor.from_tensor
tf.RaggedTensor.to_tensor
...

def __init__(self, **kwargs):
    super(RaggedToDenseTensor, self).__init__(**kwargs)

def call(self, inputs):
     if isinstance(inputs, tf.RaggedTensor):
         inputs = inputs.to_tensor()
    return inputs

[dhantule]

Hi @innat, thanks for reporting this.

There is partial support for ragged tensors in Keras 3. The from_tensor and to_tensor functionalities seem to work with Keras 3 in this gist, operations for which Tensorflow supports ragged should also work.

Please refer this issue #20073 for more clarity.

[innat]

Is there any temp support for other backend (jax, torch)?

[innat]

I have the following setup (written in keras 3), works in tensorflow backend. Now, I'm thinking how to extend it to support torch and jax.

class XYZ(Layer):
    def __init__(
        self, **kwargs
    ):
        super().__init__(**kwargs)
        ...

    def call(self, inputs):

        if isinstance(inputs, tf.RaggedTensor):
            inputs = inputs.to_tensor()
        
        batch_size = ops.shape(inputs)[0]
        clips_list = []

        ...
        ...
        return <...>

  
model = VideoSwinT(
    ...
)
umodel = keras.Sequential(
    [
        keras.layers.InputLayer(shape=(32, 224, 224, 3)),
        XYZ
        model
    ]
)

[google-ml-butler]

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

[google-ml-butler]

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

[hertschuh]

@innat ,

The feature has been added, you can use it tomorrow if you install keras-nightly.

It works like this:

x = keras.ops.convert_to_tensor(x, ragged=False)

This will densify any ragged tensor.