[TF] Updates to TensorArrayProtocol derived conformances

[eaplatanios]

@rxwei These updates enable output tensor arrays for raw ops.

Friend PR: tensorflow/swift-bindings#26 .

[eaplatanios]

Btw, I have verified that this passes all tests on my machine.

[dan-zheng]

LGTM, awesome!

[eaplatanios]

I have addressed all the comments in the last commit. :)

[dan-zheng]

I think the best way to test/merge this PR is to merge tensorflow/swift-bindings#26 first, then update checkout for swift-bindings in this PR.

cc @rxwei for your thoughts

[eaplatanios]

@dan-zheng I just made tensorflow/swift-bindings#26 backwards compatible (it can uses tfop mode when possible and fall back to eager otherwise -- e.g., for output tensor lists). This is done through a new tfop-eager-fallback mode. So, it should be possible to merge it without anything breaking. However, this can also be merged independently as it also doesn't break anything.

I also finally have a working version of a split to TensorFlowCore and TensorFlow that I plan to push soon as part of #24161 . It also allows us to compile swift-apis (which now become part of the TensorFlow package) independently with no errors. :) I'll post more details soon when I push the changes, but in this case swift-bindings becomes part of TensorFlowCore.

[eaplatanios]

Actually, please don't merge this yet. I'm working on a couple of additions that make the raw ops generation a bit more flexible.

[rxwei]

The next step is to make sure TensorFlow can get distributed as part of Swift for TensorFlow toolchain builds using build-toolchain-tensorflow. Have you explored whether that's possible?

[dan-zheng]

The next step is to make sure TensorFlow can get distributed as part of Swift for TensorFlow toolchain builds using build-toolchain-tensorflow. Have you explored whether that's possible?

I'm happy to verify build-toolchain-tensorflow works - let me know when it's ready to be tested.

[eaplatanios]

I just added a _typeList instance property to TensorArrayProtocol. This is useful for raw ops such as this:

/// Creates a dataset that emits each dim-0 slice of `components` once.
@inlinable @inline(__always)
public static func tensorSliceDataset<ToutputTypes: TensorArrayProtocol>(
  components: ToutputTypes,
  outputShapes: [TensorShape?]
) -> VariantHandle {
  let ret: VariantHandle = #tfop("TensorSliceDataset",
    components,
    Toutput_types$dtype: components._typeList,
    output_shapes: outputShapes)
  return ret
}

Not having this instance property would force us to have ToutputTypes: TensorGroup and then use Toutput_types$dtype: ToutputTypes._typeList. This is too constraining and would not work, for example, when building a dataset out of an array of tensors. With this change this works fine.

I have verified that all tests pass on my machine.

I also made the corresponding edits to tensorflow/swift-bindings#26, along with some changes to properly handle input tensor lists when using eager mode.

[eaplatanios]

All comments have been addressed.

[rxwei]

@swift-ci please test tensorflow

[rxwei]

@swift-ci please test tensorflow

[rxwei]

@swift-ci please test tensorflow

[rxwei]

@swift-ci please test tensorflow

[rxwei]

@swift-ci please test tensorflow Linux

[rxwei]

@swift-ci please test tensorflow Linux

[rxwei]

@swift-ci please test tensorflow Linux

[rxwei]

@swift-ci please test tensorflow Linux

[eaplatanios]

@rxwei not sure why this fails on the CI server. All tests pass locally for me.

[eaplatanios]

@rxwei Could we retry running the tests now that I reverted the swift-bindings changes?

[rxwei]

@swift-ci please clean test tensorflow

[rxwei]

@swift-ci please clean test tensorflow

[eaplatanios]

@rxwei I merged #24340 here so now all tests should pass (the only failed tests were the GPU strided slice assignment ones.

[rxwei]

@swift-ci please test tensorflow

[rxwei]

@swift-ci please test tensorflow

[rxwei]

@swift-ci please test tensorflow

[rxwei]

@swift-ci please test tensorflow

[rxwei]

@swift-ci please test tensorflow macOS

[rxwei]

@swift-ci please test tensorflow macOS