Please make sure that this is a feature request. As per our GitHub Policy, we only address code/doc bugs, performance issues, feature requests and build/installation issues on GitHub. tag:feature_template

System information

• TensorFlow version (you are using): TensorFlow-gpu 1.14+
• Are you willing to contribute it (Yes/No): No.
  Not this stage.

Describe the feature and the current behavior/state.
As tf.keras.model could take a raggedtensor as an input, there is a need to convert a TensorArray to a RaggedTensor.

Will this change the current api? How?
Yes, it requires a new api to TensorArray (say, to_ragged).

Who will benefit with this feature?
This allows the output (no rectangle tensors of same shape) of tf.map_fn of batch-based slicing operation as an input to a tf.keras.layers.Conv2D seamlessly.

Any Other info.
No.

Thank you