numrs2 0.3.1

A Rust implementation inspired by NumPy for numerical computing (NumRS2)
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213

//! Graph Neural Network Graph Classification Example
//!
//! This example demonstrates graph-level classification using Graph Isomorphism Networks (GIN).
//! We classify molecular graphs to predict chemical properties.
//!
//! # Task
//!
//! Given molecular graphs where:
//! - Nodes represent atoms
//! - Edges represent chemical bonds
//! - Node features represent atom properties (atomic number, charge, etc.)
//! - Goal: Classify molecules as toxic/non-toxic
//!
//! # Method
//!
//! We use GIN with graph pooling:
//! 1. GIN layers to learn node representations
//! 2. Global pooling to get graph-level representation
//! 3. MLP classifier for final prediction
//!
//! # Reference
//!
//! Xu et al. (2019) - "How Powerful are Graph Neural Networks?" (ICLR)

use numrs2::new_modules::nn::graph::*;
use scirs2_core::ndarray::Array2;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    println!("=== GIN Graph Classification Example ===\n");

    // Create two example molecular graphs

    // Molecule 1: Small non-toxic molecule (e.g., water-like)
    println!("Molecule 1: Small non-toxic structure");
    let mol1_nodes = 3; // 3 atoms
    let mol1_edges = vec![(0, 1), (1, 2)]; // linear structure
    let mol1_features = Array2::from_shape_fn((mol1_nodes, 8), |(i, j)| {
        // Simple atom features
        ((i * 3 + j) % 10) as f64 / 10.0
    });
    println!("  Nodes: {}, Edges: {}", mol1_nodes, mol1_edges.len());

    // Molecule 2: Larger toxic molecule (e.g., benzene-like ring)
    println!("Molecule 2: Larger toxic structure");
    let mol2_nodes = 6; // 6 atoms
    let mol2_edges = vec![
        (0, 1),
        (1, 2),
        (2, 3),
        (3, 4),
        (4, 5),
        (5, 0), // ring structure
    ];
    let mol2_features = Array2::from_shape_fn((mol2_nodes, 8), |(i, j)| {
        // Different atom features
        ((i * 5 + j + 3) % 10) as f64 / 10.0 + 0.3
    });
    println!("  Nodes: {}, Edges: {}", mol2_nodes, mol2_edges.len());
    println!();

    // Build GIN model
    let input_dim = 8;
    let hidden_dim = 16;
    let epsilon = 0.0; // fixed epsilon (non-learnable)

    println!("Building GIN model:");
    println!("  Input features: {}", input_dim);
    println!("  Hidden dimension: {}", hidden_dim);
    println!("  Epsilon: {} (fixed)", epsilon);
    println!("  Aggregation: SUM (most expressive for WL test)");
    println!();

    let gin_layer = GinLayer::new(input_dim, hidden_dim, epsilon)?;

    // Process Molecule 1
    println!("Processing Molecule 1...");
    let mol1_adj = SparseAdjacency::from_edges(mol1_nodes, &mol1_edges)?;
    let mol1_hidden = gin_layer.forward(&mol1_adj, &mol1_features.view())?;
    println!("  Node representations shape: {:?}", mol1_hidden.shape());

    // Apply ReLU activation
    let mut mol1_hidden_relu = mol1_hidden.clone();
    for i in 0..mol1_hidden_relu.nrows() {
        for j in 0..mol1_hidden_relu.ncols() {
            if mol1_hidden_relu[[i, j]] < 0.0 {
                mol1_hidden_relu[[i, j]] = 0.0;
            }
        }
    }

    // Global pooling to get graph representation
    let mol1_graph_repr = global_mean_pool(&mol1_hidden_relu.view())?;
    println!("  After global pooling: {:?}", mol1_graph_repr.shape());
    println!(
        "  Graph representation (first 5 dims): [{:.3}, {:.3}, {:.3}, {:.3}, {:.3}]",
        mol1_graph_repr[0],
        mol1_graph_repr[1],
        mol1_graph_repr[2],
        mol1_graph_repr[3],
        mol1_graph_repr[4]
    );

    // Process Molecule 2
    println!("\nProcessing Molecule 2...");
    let mol2_adj = SparseAdjacency::from_edges(mol2_nodes, &mol2_edges)?;
    let mol2_hidden = gin_layer.forward(&mol2_adj, &mol2_features.view())?;
    println!("  Node representations shape: {:?}", mol2_hidden.shape());

    // Apply ReLU
    let mut mol2_hidden_relu = mol2_hidden.clone();
    for i in 0..mol2_hidden_relu.nrows() {
        for j in 0..mol2_hidden_relu.ncols() {
            if mol2_hidden_relu[[i, j]] < 0.0 {
                mol2_hidden_relu[[i, j]] = 0.0;
            }
        }
    }

    // Global pooling
    let mol2_graph_repr = global_mean_pool(&mol2_hidden_relu.view())?;
    println!("  After global pooling: {:?}", mol2_graph_repr.shape());
    println!(
        "  Graph representation (first 5 dims): [{:.3}, {:.3}, {:.3}, {:.3}, {:.3}]",
        mol2_graph_repr[0],
        mol2_graph_repr[1],
        mol2_graph_repr[2],
        mol2_graph_repr[3],
        mol2_graph_repr[4]
    );

    // Compare graph representations
    println!("\nGraph Representation Comparison:");
    let mut l2_dist = 0.0;
    for i in 0..hidden_dim {
        let diff = mol1_graph_repr[i] - mol2_graph_repr[i];
        l2_dist += diff * diff;
    }
    l2_dist = l2_dist.sqrt();
    println!("  L2 distance between representations: {:.4}", l2_dist);
    println!("  Different structures produce different embeddings!");

    // Demonstrate different pooling strategies
    println!("\nPooling Strategy Comparison:");

    // Mean pooling (already computed)
    println!("  Mean pooling:");
    println!(
        "    Mol1: {:.4}",
        mol1_graph_repr.iter().sum::<f64>() / hidden_dim as f64
    );
    println!(
        "    Mol2: {:.4}",
        mol2_graph_repr.iter().sum::<f64>() / hidden_dim as f64
    );

    // Max pooling
    let mol1_max_pool = global_max_pool(&mol1_hidden_relu.view())?;
    let mol2_max_pool = global_max_pool(&mol2_hidden_relu.view())?;
    println!("  Max pooling:");
    println!(
        "    Mol1: {:.4}",
        mol1_max_pool
            .iter()
            .cloned()
            .fold(f64::NEG_INFINITY, f64::max)
    );
    println!(
        "    Mol2: {:.4}",
        mol2_max_pool
            .iter()
            .cloned()
            .fold(f64::NEG_INFINITY, f64::max)
    );

    // Sum pooling
    let mol1_sum_pool = global_sum_pool(&mol1_hidden_relu.view())?;
    let mol2_sum_pool = global_sum_pool(&mol2_hidden_relu.view())?;
    println!("  Sum pooling:");
    println!("    Mol1: {:.4}", mol1_sum_pool.iter().sum::<f64>());
    println!("    Mol2: {:.4}", mol2_sum_pool.iter().sum::<f64>());

    // Demonstrate GIN's expressiveness
    println!("\nGIN Expressiveness:");
    println!("  GIN with sum aggregation is as powerful as WL test");
    println!("  It can distinguish non-isomorphic graphs");
    println!("  Epsilon parameter: (1+ε)·h_v + Σh_u");
    println!("    ε=0 (fixed): Standard GIN");
    println!("    ε>0 (learnable): Can adapt to specific tasks");

    // Show structural information captured
    println!("\nStructural Information:");
    println!("  Molecule 1 (linear):");
    println!(
        "    Average degree: {:.2}",
        mol1_edges.len() as f64 * 2.0 / mol1_nodes as f64
    );
    println!("    Structure: Open chain");
    println!("  Molecule 2 (ring):");
    println!(
        "    Average degree: {:.2}",
        mol2_edges.len() as f64 * 2.0 / mol2_nodes as f64
    );
    println!("    Structure: Closed cycle");

    println!("\n=== Example completed successfully! ===");
    println!("\nNext steps:");
    println!("  1. Stack multiple GIN layers for deeper networks");
    println!("  2. Add edge features for bond types");
    println!("  3. Train with labeled molecular datasets");
    println!("  4. Use different pooling strategies (DiffPool, TopK)");

    Ok(())
}