Add decode_json support

[yongtang]

This PR adds decode_json support so that it is possible to use

values = tfio.experimental.serialization.decode_json(json, specs)

to decode a JSON string into a structured Tensors.

The specs field is the structured tf.TensorSpec that is necessary
to reconstruct the JSON values. The reason being JSON could be intepreted
differently in lists:

{
"v": [1, 2, 3]
}

could be considered as one Tensor of shape (3), or a 3 Tensor structure within dict "v".

Note structured tf.TensorSpecs need to change list [] into tuple ()
as list is not supported in structures in TensorFlow.

This PR fixes #692.

Signed-off-by: Yong Tang yong.tang.github@outlook.com

[yongtang]

/cc @noahtren

[noahtren]

Thanks so much! This is a huge help to my pipeline which includes JSON alongside a lot of other data types. So helpful to have a function to turn these into tensors.

[djl11]

The following example does not parse the json correctly:

core_ops.io_decode_json(tf.constant('{"x":[[1.0]]}'), ['/x'], [tf.float32])

Gives:

[<tf.Tensor: shape=(1,), dtype=float32, numpy=array([4.2949673e+09], dtype=float32)>]

However, 1-D works, for example:

core_ops.io_decode_json(tf.constant('{"x":[1.0, 2.0]}'), ['/x'], [tf.float32])

Gives:

[<tf.Tensor: shape=(2,), dtype=float32, numpy=array([1., 2.], dtype=float32)>]

Is this desired behaviour?
I am not sure how to load a multi-dimensional tensor stored in a json file as a nested list.

[yongtang]

@djl11 A PR #1101 has been opened to support the 2-D tensor in your example.

[djl11]

I think with recursion the multi-dimensional case shouldn't be too hard, unless I'm missing something. If I find the time I'll try and implement this myself. Thanks for putting in the new PR @yongtang!