How do I specify a ragged tensor as an input object in a sequential model (Keras v3)

[fninsiima]

I am following the object detection tutorial with yolov8 and kerascv

However, instead of using yolo for object detection, I would like to use a custom sequential model.

My roadblock is at the point of defining the input layer. I would like it to take in a ragged tensor.

According to the keras v2 docs, there is an option to specify ragged=True and define the input layer via inputs = tf.keras.Input(shape = [] , dtype = tf.int64, ragged = True)

However in keras v3, the ragged option has been removed.

So how do I proceed with defining an input layer that takes in a ragged tensor?

[fchollet]

There is no ragged tensor support in Keras 3 at this time. Ragged tensors are a TF-only concept (it doesn't exist in JAX or PyTorch) and it is incompatible with XLA, which is the default compiler when using the TF backend in Keras 3.

In general my recommendation is that you can have ragged tensors in your input preprocessing pipeline, but by the time you hit the model you need to densify / pad your tensors.