Getting error when serialize a tfds object

What I need help with / What I was wondering
I have tried serialize the response object from this code:

import tensorflow_datasets as tfds
(ds_train, ds_test), ds_info = tfds.load(
    'mnist',
    split=['train', 'test'],
    shuffle_files=True,
    as_supervised=True,
    with_info=True,
)
response = {
  "ds_train": ds_train,
  "ds_test": ds_test,
  "ds_info": ds_info,
}

But when i use the pickle.dump method to convert this dict to a binary file this exception is throwed:
tensorflow.python.framework.errors_impl.InternalError: Tensorflow type 21 not convertible to numpy dtype.

Can someone help me?

Environment information
(if applicable)

• Operating System: Debian
• Python version: 3.7
• tensorflow-datasets/tfds-nightly version: stable

Reproduce the pickle.dump failure with the shown MNIST tfds.load example under Python 3.7, Debian, and the stable tensorflow-datasets version. No repository file, entry point, or test is named; done means determining whether these response objects are supported for serialization and capturing the result in a focused regression check or documented resolution.