few_shot_learning/
few_shot_learning.rs

1#![allow(clippy::pedantic, clippy::unnecessary_wraps)]
2//! Quantum Few-Shot Learning Example
3//!
4//! This example demonstrates how to use quantum few-shot learning algorithms
5//! to learn from very limited training examples.
6
7use quantrs2_ml::autodiff::optimizers::Adam;
8use quantrs2_ml::prelude::*;
9use quantrs2_ml::qnn::QNNLayerType;
10use scirs2_core::ndarray::{Array1, Array2};
11use scirs2_core::random::prelude::*;
12
13fn main() -> Result<()> {
14    println!("=== Quantum Few-Shot Learning Demo ===\n");
15
16    // Step 1: Generate synthetic dataset
17    println!("1. Generating synthetic dataset for 5-way classification...");
18    let num_samples_per_class = 20;
19    let num_classes = 5;
20    let num_features = 4;
21    let total_samples = num_samples_per_class * num_classes;
22
23    // Generate data with different patterns for each class
24    let mut data = Array2::zeros((total_samples, num_features));
25    let mut labels = Array1::zeros(total_samples);
26
27    for class_id in 0..num_classes {
28        for sample_idx in 0..num_samples_per_class {
29            let idx = class_id * num_samples_per_class + sample_idx;
30
31            // Create class-specific patterns
32            for feat in 0..num_features {
33                data[[idx, feat]] = 0.1f64.mul_add(
34                    2.0f64.mul_add(thread_rng().gen::<f64>(), -1.0),
35                    (sample_idx as f64)
36                        .mul_add(0.1, (class_id as f64).mul_add(0.5, feat as f64 * 0.3))
37                        .sin(),
38                );
39            }
40            labels[idx] = class_id;
41        }
42    }
43
44    println!(
45        "   Dataset created: {total_samples} samples, {num_features} features, {num_classes} classes"
46    );
47
48    // Step 2: Create quantum model for few-shot learning
49    println!("\n2. Creating quantum neural network...");
50    let layers = vec![
51        QNNLayerType::EncodingLayer { num_features },
52        QNNLayerType::VariationalLayer { num_params: 8 },
53        QNNLayerType::EntanglementLayer {
54            connectivity: "circular".to_string(),
55        },
56        QNNLayerType::VariationalLayer { num_params: 8 },
57        QNNLayerType::MeasurementLayer {
58            measurement_basis: "computational".to_string(),
59        },
60    ];
61
62    let qnn = QuantumNeuralNetwork::new(layers, 4, num_features, num_classes)?;
63    println!("   Quantum model created with {} qubits", qnn.num_qubits);
64
65    // Step 3: Test different few-shot learning methods
66    println!("\n3. Testing few-shot learning methods:");
67
68    // Method 1: Prototypical Networks
69    println!("\n   a) Prototypical Networks (5-way 3-shot):");
70    test_prototypical_networks(&data, &labels, qnn.clone())?;
71
72    // Method 2: MAML
73    println!("\n   b) Model-Agnostic Meta-Learning (MAML):");
74    test_maml(&data, &labels, qnn.clone())?;
75
76    // Step 4: Compare performance across different shot values
77    println!("\n4. Performance comparison across different K-shot values:");
78    compare_shot_performance(&data, &labels, qnn)?;
79
80    println!("\n=== Few-Shot Learning Demo Complete ===");
81
82    Ok(())
83}
84
85/// Test prototypical networks
86fn test_prototypical_networks(
87    data: &Array2<f64>,
88    labels: &Array1<usize>,
89    qnn: QuantumNeuralNetwork,
90) -> Result<()> {
91    let mut learner = FewShotLearner::new(FewShotMethod::PrototypicalNetworks, qnn);
92
93    // Generate episodes for training
94    let num_episodes = 10;
95    let mut episodes = Vec::new();
96
97    for _ in 0..num_episodes {
98        let episode = FewShotLearner::generate_episode(
99            data, labels, 5, // 5-way
100            3, // 3-shot
101            5, // 5 query examples per class
102        )?;
103        episodes.push(episode);
104    }
105
106    // Train
107    let mut optimizer = Adam::new(0.01);
108    let accuracies = learner.train(&episodes, &mut optimizer, 20)?;
109
110    // Print results
111    println!("   Training completed:");
112    println!("   - Initial accuracy: {:.2}%", accuracies[0] * 100.0);
113    println!(
114        "   - Final accuracy: {:.2}%",
115        accuracies.last().unwrap() * 100.0
116    );
117    println!(
118        "   - Improvement: {:.2}%",
119        (accuracies.last().unwrap() - accuracies[0]) * 100.0
120    );
121
122    Ok(())
123}
124
125/// Test MAML
126fn test_maml(data: &Array2<f64>, labels: &Array1<usize>, qnn: QuantumNeuralNetwork) -> Result<()> {
127    let mut learner = FewShotLearner::new(
128        FewShotMethod::MAML {
129            inner_steps: 5,
130            inner_lr: 0.01,
131        },
132        qnn,
133    );
134
135    // Generate meta-training tasks
136    let num_tasks = 20;
137    let mut tasks = Vec::new();
138
139    for _ in 0..num_tasks {
140        let task = FewShotLearner::generate_episode(
141            data, labels, 3, // 3-way (fewer classes for MAML)
142            5, // 5-shot
143            5, // 5 query examples
144        )?;
145        tasks.push(task);
146    }
147
148    // Meta-train
149    let mut meta_optimizer = Adam::new(0.001);
150    let losses = learner.train(&tasks, &mut meta_optimizer, 10)?;
151
152    println!("   Meta-training completed:");
153    println!("   - Initial loss: {:.4}", losses[0]);
154    println!("   - Final loss: {:.4}", losses.last().unwrap());
155    println!(
156        "   - Convergence rate: {:.2}%",
157        (1.0 - losses.last().unwrap() / losses[0]) * 100.0
158    );
159
160    Ok(())
161}
162
163/// Compare performance across different K-shot values
164fn compare_shot_performance(
165    data: &Array2<f64>,
166    labels: &Array1<usize>,
167    qnn: QuantumNeuralNetwork,
168) -> Result<()> {
169    let k_values = vec![1, 3, 5, 10];
170
171    for k in k_values {
172        println!("\n   Testing {k}-shot learning:");
173
174        let mut learner = FewShotLearner::new(FewShotMethod::PrototypicalNetworks, qnn.clone());
175
176        // Generate episodes
177        let mut episodes = Vec::new();
178        for _ in 0..5 {
179            let episode = FewShotLearner::generate_episode(
180                data, labels, 3, // 3-way
181                k, // k-shot
182                5, // 5 query
183            )?;
184            episodes.push(episode);
185        }
186
187        // Quick training
188        let mut optimizer = Adam::new(0.01);
189        let accuracies = learner.train(&episodes, &mut optimizer, 10)?;
190
191        println!(
192            "     Final accuracy: {:.2}%",
193            accuracies.last().unwrap() * 100.0
194        );
195    }
196
197    Ok(())
198}
199
200/// Demonstrate episode structure
201fn demonstrate_episode_structure() -> Result<()> {
202    println!("\n5. Episode Structure Demonstration:");
203
204    // Create a simple episode manually
205    let support_set = vec![
206        // Class 0
207        (Array1::from_vec(vec![0.1, 0.2, 0.3, 0.4]), 0),
208        (Array1::from_vec(vec![0.15, 0.25, 0.35, 0.45]), 0),
209        // Class 1
210        (Array1::from_vec(vec![0.8, 0.7, 0.6, 0.5]), 1),
211        (Array1::from_vec(vec![0.85, 0.75, 0.65, 0.55]), 1),
212    ];
213
214    let query_set = vec![
215        (Array1::from_vec(vec![0.12, 0.22, 0.32, 0.42]), 0),
216        (Array1::from_vec(vec![0.82, 0.72, 0.62, 0.52]), 1),
217    ];
218
219    let episode = Episode {
220        support_set,
221        query_set,
222        num_classes: 2,
223        k_shot: 2,
224    };
225
226    println!("   2-way 2-shot episode created");
227    println!("   - Support set size: {}", episode.support_set.len());
228    println!("   - Query set size: {}", episode.query_set.len());
229
230    Ok(())
231}
few_shot_learning/few_shot_learning.rs

few_shot_learning/
few_shot_learning.rs
