mirror of
https://github.com/ruvnet/RuVector.git
synced 2026-05-23 12:55:26 +00:00
1cc09933eb
This commit addresses GitHub issue #17 by implementing comprehensive forgetting mitigation for continual learning in the GNN module. ## New Features ### Optimizer Implementation (training.rs) - Full Adam optimizer with bias-corrected first and second moments - SGD with momentum support - Lazy initialization of state buffers for efficiency ### Replay Buffer (replay.rs) - Experience replay with reservoir sampling for uniform distribution - Distribution shift detection with statistical tracking - Configurable capacity and batch sampling ### Elastic Weight Consolidation (ewc.rs) - Fisher information diagonal computation - Anchor weight consolidation for task boundaries - EWC penalty and gradient computation ### Learning Rate Scheduling (scheduler.rs) - Constant, StepDecay, Exponential schedulers - CosineAnnealing with warm restarts - WarmupLinear for pre-training warmup - ReduceOnPlateau for adaptive learning ## Deployment Infrastructure ### GitHub Actions Release Pipeline (.github/workflows/release.yml) - 8-stage CI/CD pipeline for complete releases - Validates, builds crates, WASM, and native modules - Publishes to crates.io and npmjs.com - Creates GitHub releases with artifacts ### Deployment Script (scripts/deploy.sh) - Comprehensive deployment orchestration - Version synchronization across Cargo.toml and package.json - Dry-run mode for testing - Cross-platform native builds support ## Test Coverage - 177 tests passing in ruvector-gnn - Comprehensive tests for all new modules - Convergence tests for optimizers 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| deploy-20251126-225900.log | ||
| deploy-20251126-230017.log | ||