Richard Wei

#111 added derivatives for some `Complex` APIs but the implementation did not consider `.infinity`, `.nan`, overflow and underflow. The following comments made by Steve should be addressed at some point:...

Guidelines for Python code examples in derivative tests.

2

What has really helped in assuring our layers' correctness is adding a Python reproducer per layer test like [this example](https://github.com/tensorflow/swift-apis/blob/9616543321a279ebb550fbfdd37fb10d63842f25/Tests/TensorFlowTests/LayerTests.swift#L310). Thank you @eaplatanios, @Shashi456, @jon-tow and @t-ae for doing this!...

Swifty API discovery via '@available' on raw APIs.

8

For many raw APIs, we already have corresponding idiomatic APIs that implement all of the raw API's functionality, e.g. `Tensor.init(_:)` for `Raw.cast(_:)`, `Tensor.reshaped(to:)` for `Raw.reshape(_:shape:)`, etc. These raw APIs are...

Doc comments for non-mutating methods often start with `Returns a ... by ing ...` instead of `s ...`. Quite a few doc comments in the library do not follow this...

Provide a way to access non-differentiable model state from `Optimizer.update`.

6

@dominikgrewe pointed out that K-FAC needs to access the model's non-differentiable internal state from `Optimizer.update(_:along:)`. My initial idea is to change the optimizer API to the following: ```swift public protocol...

Implement staging

2

Implement a staging back-end code generator / interpreter as a seamless plugin to support the current tagless front-end. Contributions are welcome.

[AutoDiff] WIP: Use owned callee convention for linear maps.

42

Switch to `@callee_owned` callee convention for all linear map functions (differentials and pullbacks) returned from derivative functions. This reduces a half of reference counting operations in compiler-generated derivatives, and enables...