SKaiNET

SKaiNET is an open-source deep learning framework written in Kotlin, designed with developers in mind to enable the creation modern AI powered applications with ease.

Quick example: training/eval phases with context wrappers

SKaiNET provides small helper scopes to override the execution phase for a given call without mutating your base context:

• train(ctx) { ... } — runs the block with phase = TRAIN
• eval(ctx) { ... } — runs the block with phase = EVAL

This is useful for modules that behave differently in training vs evaluation (e.g., Dropout, BatchNorm).

Kotlin

import sk.ainet.context.train import sk.ainet.context.eval import sk.ainet.lang.nn.DefaultNeuralNetworkExecutionContext // import your model class that defines forward(x, ctx)

val base = DefaultNeuralNetworkExecutionContext() // default phase is EVAL val x = /* prepare your input tensor / val model = / construct your model */

// Force TRAIN phase for this call only val yTrain = train(base) { ctx -> model.forward(x, ctx) }

// Force EVAL phase for this call only val yEval = eval(base) { ctx -> model.forward(x, ctx) }

// You can still call with the base context directly (uses its phase) val yBase = model.forward(x, base)

Development Practices

This project follows established development practices for maintaining code quality and release management:

• Branching Model: We use GitFlow as our branching strategy for managing feature development, releases, and hotfixes.
• Versioning: We follow Semantic Versioning (SemVer) for all releases, ensuring predictable version numbering based on the nature of changes.

Reflective Documentation (short overview)

SKaiNET includes a reflective documentation system that keeps docs in sync with the code. During the build, a KSP processor extracts operator metadata (signatures, parameters, backend availability, implementation status) into a JSON file. A small DocGen tool then converts this JSON into AsciiDoc fragments and pages.

• Source of truth (generated): skainet-lang/skainet-lang-core/build/generated/ksp/metadata/commonMain/resources/operators.json
• Generated docs output: docs/modules/operators/generated/
• Asciidoctor site output: build/docs/asciidoc/ (if you run an Asciidoctor task locally)

Quick start: generate reflective docs

Use any of the following Gradle tasks from the project root:

1. Full pipeline (recommended) ./gradlew generateDocs

   • Runs KSP to produce operators.json (if needed)
   • Generates AsciiDoc files under docs/modules/operators/generated
   • Optionally, you can run an Asciidoctor task to build an HTML site locally (output under build/docs/asciidoc)

2. Operators documentation only ./gradlew generateOperatorDocs

   • Depends on KSP; runs the built-in generateDocs task and then Asciidoctor

Open the generated AsciiDoc sources in docs/modules/operators/generated with your preferred AsciiDoc viewer. If you build an HTML site locally with Asciidoctor, open build/docs/asciidoc.

Documentation tooling

We now use the Gradle plugin (buildSrc) only. The former skainet-lang-export-ops module has been removed. All everyday workflows are covered by:

• generateDocs — converts KSP JSON to AsciiDoc
• validateOperatorSchema — validates generated operators.json against the JSON schema

Run from the project root, for example:

• ./gradlew generateDocs
• ./gradlew validateOperatorSchema

---

Development Practices

This project follows established development practices for maintaining code quality and release management:

• Branching Model: We use GitFlow as our branching strategy for managing feature development, releases, and hotfixes.
• Versioning: We follow Semantic Versioning (SemVer) for all releases, ensuring predictable version numbering based on the nature of changes.