[dan-zheng]

Add Sequential{N} generic typealiases.
These are easier to use than long nested Sequential types.

import TensorFlow
// Before:
typealias Model = Sequential<Dense<Float>, Sequential<Dense<Float>, Dense<Float>>>
// After:
typealias Model = Sequential3<Dense<Float>, Dense<Float>, Dense<Float>>

[dan-zheng]

Generic typealiases with additional constraints aren't supported, so we can't land this anytime soon. I asked about potential support on the Swift forums.

---

In the meantime, we'll have to deal with long nested Sequential types:

var module: Sequential<Sequential<Conv2D<Float>, Sequential<Function<Tensor<Float>, Tensor<Float>>, Sequential<Conv2D<Float>, Sequential<BatchNorm<Float>, Sequential<Function<Tensor<Float>, Tensor<Float>>, Sequential<Conv2D<Float>, Sequential<BatchNorm<Float>, Function<Tensor<Float>, Tensor<Float>>>>>>>>>, Sequential<ConvLayer, Sequential<BatchNorm<Float>, Sequential<Function<Tensor<Float>, Tensor<Float>>, ConvLayer>>>>

An alternative to long nested Sequential types is type erasure:

• A manual approach like AnyLayer.
• Or eventually via language features like opaque type aliases.

[dan-zheng]

Hey, generic typealiases can have where clauses! So generic Sequential{N} typealiases do actually work.

Ready for review.

[saeta]

This seems like a nice improvement. Thanks!