make everything compile on the stock toolchain

marcrasi commented 3 years ago

This makes everything compile on the stock toolchain! Also everything still compiles on tensorflow-0.12 toolchain, so that you can continue working with that.

@ProfFan, let me know if it's okay for me to merge this with the tracking branch and also update the tracking branch to work with the stock toolchain. If it's okay, I'll do that next.

Note: The process for installing the stock toolchain with the required TensorFlow libraries is currently a bit complicated, so I wouldn't suggest trying it yourself yet. Soon we should have a simplified process. But if you really want to try it yourself, let me know and I can describe the process.

The required changes were:

import _Differentiation to enable differentiation
The stock toolchain seems a bit more strict about requiring @differentiable annotations on functions that are called by other @differentiable functions, so I added a few more @differentiable annotations.
Conditionally import PythonKit only if it is available, because it's not available by default on stock toolchains.
Delete the dependency on swift-models (because it doesn't compile on stock toolchains yet), and copy-paste the required code from swift-models into a ModelSupport module.

dabrahams commented 3 years ago

Nice! I get: warning: dependency 'swift-argument-parser' is not used by any target.

Suggest you remove that from Package.swift

marcrasi commented 3 years ago

Nice! I get: warning: dependency 'swift-argument-parser' is not used by any target.

Suggest you remove that from Package.swift

I added a TODO. I can't do it now because there is some bug in SwiftPM so that it fails to build without it. It's fixed in very recent toolchains, so we should be able to remove it soon.

dabrahams commented 3 years ago

How do I get a stock toolchain this works with?

borglab / SwiftFusion

make everything compile on the stock toolchain #251