172 lines
5.1 KiB
Markdown
172 lines
5.1 KiB
Markdown
# Ruvector GNN Layer Implementation
|
||
|
||
## Overview
|
||
|
||
Implemented a complete Graph Neural Network (GNN) layer for Ruvector that operates on HNSW topology, providing message passing, attention mechanisms, and recurrent state updates.
|
||
|
||
## Location
|
||
|
||
**Implementation:** `/home/user/ruvector/crates/ruvector-gnn/src/layer.rs`
|
||
|
||
## Components Implemented
|
||
|
||
### 1. Linear Layer
|
||
- **Purpose:** Weight matrix multiplication for transformations
|
||
- **Initialization:** Xavier/Glorot initialization for stable gradients
|
||
- **API:**
|
||
```rust
|
||
Linear::new(input_dim: usize, output_dim: usize) -> Self
|
||
forward(&self, input: &[f32]) -> Vec<f32>
|
||
```
|
||
|
||
### 2. Layer Normalization
|
||
- **Purpose:** Normalize activations for stable training
|
||
- **Features:** Learnable scale (gamma) and shift (beta) parameters
|
||
- **API:**
|
||
```rust
|
||
LayerNorm::new(dim: usize, eps: f32) -> Self
|
||
forward(&self, input: &[f32]) -> Vec<f32>
|
||
```
|
||
|
||
### 3. Multi-Head Attention
|
||
- **Purpose:** Attention-based neighbor aggregation
|
||
- **Features:**
|
||
- Separate Q, K, V projections
|
||
- Scaled dot-product attention
|
||
- Multi-head parallelization
|
||
- **API:**
|
||
```rust
|
||
MultiHeadAttention::new(embed_dim: usize, num_heads: usize) -> Self
|
||
forward(&self, query: &[f32], keys: &[Vec<f32>], values: &[Vec<f32>]) -> Vec<f32>
|
||
```
|
||
|
||
### 4. GRU Cell (Gated Recurrent Unit)
|
||
- **Purpose:** State updates with gating mechanisms
|
||
- **Features:**
|
||
- Update gate: Controls how much of new information to accept
|
||
- Reset gate: Controls how much of past information to forget
|
||
- Candidate state: Proposes new hidden state
|
||
- **API:**
|
||
```rust
|
||
GRUCell::new(input_dim: usize, hidden_dim: usize) -> Self
|
||
forward(&self, input: &[f32], hidden: &[f32]) -> Vec<f32>
|
||
```
|
||
|
||
### 5. RuvectorLayer (Main GNN Layer)
|
||
- **Purpose:** Complete GNN layer combining all components
|
||
- **Architecture:**
|
||
1. Message passing through linear transformations
|
||
2. Attention-based neighbor aggregation
|
||
3. Weighted message aggregation using edge weights
|
||
4. GRU-based state update
|
||
5. Dropout regularization
|
||
6. Layer normalization
|
||
- **API:**
|
||
```rust
|
||
RuvectorLayer::new(
|
||
input_dim: usize,
|
||
hidden_dim: usize,
|
||
heads: usize,
|
||
dropout: f32
|
||
) -> Self
|
||
|
||
forward(
|
||
&self,
|
||
node_embedding: &[f32],
|
||
neighbor_embeddings: &[Vec<f32>],
|
||
edge_weights: &[f32]
|
||
) -> Vec<f32>
|
||
```
|
||
|
||
## Usage Example
|
||
|
||
```rust
|
||
use ruvector_gnn::RuvectorLayer;
|
||
|
||
// Create GNN layer: 128-dim input -> 256-dim hidden, 4 attention heads, 10% dropout
|
||
let layer = RuvectorLayer::new(128, 256, 4, 0.1);
|
||
|
||
// Node and neighbor embeddings
|
||
let node = vec![0.5; 128];
|
||
let neighbors = vec![
|
||
vec![0.3; 128],
|
||
vec![0.7; 128],
|
||
];
|
||
let edge_weights = vec![0.8, 0.6]; // e.g., inverse distances
|
||
|
||
// Forward pass
|
||
let updated_embedding = layer.forward(&node, &neighbors, &edge_weights);
|
||
// Output: 256-dimensional embedding
|
||
```
|
||
|
||
## Key Features
|
||
|
||
1. **HNSW-Aware:** Designed to operate on HNSW graph topology
|
||
2. **Message Passing:** Transforms and aggregates neighbor information
|
||
3. **Attention Mechanism:** Learns importance of different neighbors
|
||
4. **Edge Weights:** Incorporates graph structure (e.g., distances)
|
||
5. **State Updates:** GRU cells maintain and update node states
|
||
6. **Normalization:** Layer norm for training stability
|
||
7. **Regularization:** Dropout to prevent overfitting
|
||
|
||
## Mathematical Operations
|
||
|
||
### Forward Pass Flow:
|
||
```
|
||
1. node_msg = W_msg × node_embedding
|
||
2. neighbor_msgs = [W_msg × neighbor_i for all neighbors]
|
||
3. attention_out = MultiHeadAttention(node_msg, neighbor_msgs)
|
||
4. weighted_msgs = Σ(weight_i × neighbor_msg_i) / Σ(weights)
|
||
5. combined = attention_out + weighted_msgs
|
||
6. aggregated = W_agg × combined
|
||
7. updated = GRU(aggregated, node_msg)
|
||
8. dropped = Dropout(updated)
|
||
9. output = LayerNorm(dropped)
|
||
```
|
||
|
||
## Testing
|
||
|
||
All components include comprehensive unit tests:
|
||
- ✓ Linear layer transformation
|
||
- ✓ Layer normalization (zero mean check)
|
||
- ✓ Multi-head attention with multiple neighbors
|
||
- ✓ GRU state updates
|
||
- ✓ RuvectorLayer with neighbors
|
||
- ✓ RuvectorLayer without neighbors (edge case)
|
||
|
||
**Test Results:** All 6 layer tests passing
|
||
|
||
## Integration
|
||
|
||
The layer integrates with existing ruvector-gnn components:
|
||
- Used in `search.rs` for hierarchical forward passes
|
||
- Compatible with HNSW topology from `ruvector-core`
|
||
- Supports differentiable search operations
|
||
|
||
## Dependencies
|
||
|
||
- **ndarray:** Matrix operations and linear algebra
|
||
- **rand/rand_distr:** Weight initialization
|
||
- **serde:** Serialization support
|
||
|
||
## Performance Considerations
|
||
|
||
1. **Xavier Initialization:** Helps gradient flow during training
|
||
2. **Batch Operations:** Uses ndarray for efficient matrix ops
|
||
3. **Attention Caching:** Could be added for repeated queries
|
||
4. **Edge Weight Normalization:** Ensures stable aggregation
|
||
|
||
## Future Enhancements
|
||
|
||
1. Actual dropout sampling (current: deterministic scaling)
|
||
2. Gradient computation for training
|
||
3. Batch processing support
|
||
4. GPU acceleration via specialized backends
|
||
5. Additional aggregation schemes (mean, max, sum)
|
||
|
||
---
|
||
|
||
**Status:** ✅ Implemented and tested successfully
|
||
**Build:** ✅ Compiles without errors (warnings: documentation only)
|
||
**Tests:** ✅ 26/26 tests passing
|