Crate ruvector_gnn 

RuVector GNN

Graph Neural Network capabilities for RuVector, providing tensor operations, GNN layers, compression, and differentiable search.

Forgetting Mitigation (Issue #17)

This crate includes comprehensive forgetting mitigation for continual learning:

• Adam Optimizer: Full implementation with momentum and bias correction
• Replay Buffer: Experience replay with reservoir sampling for uniform coverage
• EWC (Elastic Weight Consolidation): Prevents catastrophic forgetting
• Learning Rate Scheduling: Multiple strategies including warmup and plateau detection

Usage Example

use ruvector_gnn::{
    training::{Optimizer, OptimizerType},
    replay::ReplayBuffer,
    ewc::ElasticWeightConsolidation,
    scheduler::{LearningRateScheduler, SchedulerType},
};

// Create Adam optimizer
let mut optimizer = Optimizer::new(OptimizerType::Adam {
    learning_rate: 0.001,
    beta1: 0.9,
    beta2: 0.999,
    epsilon: 1e-8,
});

// Create replay buffer for experience replay
let mut replay = ReplayBuffer::new(10000);

// Create EWC for preventing forgetting
let mut ewc = ElasticWeightConsolidation::new(0.4);

// Create learning rate scheduler
let mut scheduler = LearningRateScheduler::new(
    SchedulerType::CosineAnnealing { t_max: 100, eta_min: 1e-6 },
    0.001
);

Re-exports

pub use compress::CompressedTensor;

pub use compress::CompressionLevel;

pub use compress::TensorCompress;

pub use error::GnnError;

pub use error::Result;

pub use ewc::ElasticWeightConsolidation;

pub use layer::RuvectorLayer;

pub use query::QueryMode;

pub use query::QueryResult;

pub use query::RuvectorQuery;

pub use query::SubGraph;

pub use replay::DistributionStats;

pub use replay::ReplayBuffer;

pub use replay::ReplayEntry;

pub use scheduler::LearningRateScheduler;

pub use scheduler::SchedulerType;

pub use search::cosine_similarity;

pub use search::differentiable_search;

pub use search::hierarchical_forward;

pub use training::info_nce_loss;

pub use training::local_contrastive_loss;

pub use training::sgd_step;

pub use training::Loss;

pub use training::LossType;

pub use training::OnlineConfig;

pub use training::Optimizer;

pub use training::OptimizerType;

pub use training::TrainConfig;

pub use mmap::AtomicBitmap;

pub use mmap::MmapGradientAccumulator;

pub use mmap::MmapManager;

Modules

compress

Tensor compression with adaptive level selection

error

Error types for the GNN module.

ewc

layer

GNN Layer Implementation for HNSW Topology

mmap

Memory-mapped embedding management for large-scale GNN training.

query

Query API for RuVector GNN

replay

Experience Replay Buffer for GNN Training

scheduler

Learning rate scheduling for Graph Neural Networks

search

tensor

Tensor operations for GNN computations.

training

Training utilities for GNN models.