Hi,

Right now, to build a new dataset op, you need to access to the header file tensorflow/core/kernels/dataset.h, but the tensorflow wheel does not expose this header.

Use case: I built new Dataset Ops to read Kaldi's "Table" I/O format to enable others to be able to move from Kaldi-based automatic speech recognition recipes to tensorflow-based ones without having to do a bunch of extra data-munging. Right now, I require users to build tensorflow from source code and point my build to the tensorflow source code path, so I can guarantee that I have access to a header file compatible with their binary. I'd prefer to be able to build my package by depending only on pip-installed tensorflow to make things easier on users.

I manually verified that adding

"//tensorflow/core/kernels:dataset",

to the deps of

transitive_hdrs(
    name = "included_headers",
    deps = [
        "//tensorflow/core:core_cpu",
        "//tensorflow/core:framework",
        "//tensorflow/core:lib",
        "//tensorflow/core:protos_all_cc",  
        "//tensorflow/core:stream_executor",
        "//third_party/eigen3",
    ],
)

in tensorflow/tools/pip_package/BUILD will include the right header file.

This is a pretty small change. Is there a particular reason why tensorflow does not already expose the header file? Is this an oversight or because you aren't ready to expose this interface publically?