Invalid Tensor Index 'ragged_flat_values' when using Sequence as top-level feature

[hermannsblum]

Short description I'm trying to add a new dataset. Serialization succeeded, but when loading the data, the error below occurs.

TypeError: Only integers, slices (`:`), ellipsis (`...`), tf.newaxis (`None`) and scalar tf.int32/tf.int64 tensors are valid indices, got 'ragged_flat_values'  

Detailed stacktrace below, the error happens in the deserialization.

Environment information

• Operating System: ubuntu 18.04
• Python version: 3.6
• tensorflow-datasets/tfds-nightly version: tfds-nightly
• tensorflow/tensorflow-gpu/tf-nightly/tf-nightly-gpu version: tested on tensorflow 2.2.0 and 2.3.0

Reproduction instructions

https://github.com/hermannsblum/tf_datasets/blob/3dpw/tensorflow_datasets/human_pose/pose_3dpw.py

Link to logs

stack trace

Expected behavior The dataset to load, or an error at dataset building.

Additional context Add any other context about the problem here.

[hermannsblum]

The cause of the bug turns out to be passing a Sequence instead of a FeaturesDict as the top level feature to DatasetInfo. Therefore:

• if features should always be encapsulated in a dict, this should be mentioned in the guide.
• otherwise, this is a bug that should be fixed.

Can a maintainer shed light on which of these options is correct?

[Conchylicultor]

Sequence({}) should be allowed at top-level feature, so this is likely a bug. Thank you for reporting

[kyamagu]

Sequence(Scalar()) works, but Sequence(Sequence(Scalar(), length=4)) fails.