Preprints and final documentation updates

[Brad Larson]

Brennan announced these on Twitter, but for those who don't follow us there: three new preprints of papers related to the Swift for TensorFlow project have been posted to arXiv:

"Swift for TensorFlow: A portable, flexible platform for deep learning"

- Accepted to MLSys 2021, this provides an overview of the project and some of the core areas we explored. Don't miss some of the neat new benchmarks towards the end, as well as our community acknowledgments.

"LazyTensor: combining eager execution with domain-specific compilers"

- A joint publication with the PyTorch XLA team, this describes the mechanism at the core of PyTorch XLA and the Swift for TensorFlow X10 backend.

"Tensors Fitting Perfectly"

- A detailed description of Adam Paszke's exploration of static analysis to determine shape mismatches in Swift models, and to even fill in the gaps for certain underspecified shapes.

In addition to these, we've added a last series of guides to the Swift for TensorFlow overview documentation at tensorflow.org/swift. These guides include:

• Python interoperability
• Tensor overview
• Accelerator backends
• Debugging X10 issues
• Layers and optimizers
• Datasets
• Training loop
• Model checkpoints
• Model summaries

Hopefully, these new preprints and guides capture a lot of what we've learned as a team and community while building Swift for TensorFlow. Thank you again to everyone who has worked with us, provided design suggestions, or just taken an interest in Swift for TensorFlow.