transfer_learning/
transfer_learning.rs

1#![allow(
2    clippy::pedantic,
3    clippy::unnecessary_wraps,
4    clippy::needless_range_loop,
5    clippy::useless_vec,
6    clippy::needless_collect,
7    clippy::too_many_arguments
8)]
9//! Quantum Transfer Learning Example
10//!
11//! This example demonstrates how to use pre-trained quantum models
12//! and fine-tune them for new tasks with transfer learning.
13
14use quantrs2_ml::autodiff::optimizers::Adam;
15use quantrs2_ml::prelude::*;
16use quantrs2_ml::qnn::QNNLayerType;
17use scirs2_core::ndarray::{Array1, Array2};
18
19fn main() -> Result<()> {
20    println!("=== Quantum Transfer Learning Demo ===\n");
21
22    // Step 1: Load a pre-trained model from the model zoo
23    println!("1. Loading pre-trained image classifier...");
24    let pretrained = QuantumModelZoo::get_image_classifier()?;
25
26    println!("   Pre-trained model info:");
27    println!("   - Task: {}", pretrained.task_description);
28    println!(
29        "   - Original accuracy: {:.2}%",
30        pretrained
31            .performance_metrics
32            .get("accuracy")
33            .unwrap_or(&0.0)
34            * 100.0
35    );
36    println!("   - Number of qubits: {}", pretrained.qnn.num_qubits);
37
38    // Step 2: Create new layers for the target task
39    println!("\n2. Creating new layers for text classification task...");
40    let new_layers = vec![
41        QNNLayerType::VariationalLayer { num_params: 6 },
42        QNNLayerType::MeasurementLayer {
43            measurement_basis: "Pauli-Z".to_string(),
44        },
45    ];
46
47    // Step 3: Initialize transfer learning with different strategies
48    println!("\n3. Testing different transfer learning strategies:");
49
50    // Strategy 1: Fine-tuning
51    println!("\n   a) Fine-tuning strategy (train last 2 layers only)");
52    let mut transfer_finetune = QuantumTransferLearning::new(
53        pretrained.clone(),
54        new_layers.clone(),
55        TransferStrategy::FineTuning {
56            num_trainable_layers: 2,
57        },
58    )?;
59
60    // Strategy 2: Feature extraction
61    println!("   b) Feature extraction strategy (freeze all pre-trained layers)");
62    let transfer_feature = QuantumTransferLearning::new(
63        pretrained.clone(),
64        new_layers.clone(),
65        TransferStrategy::FeatureExtraction,
66    )?;
67
68    // Strategy 3: Progressive unfreezing
69    println!("   c) Progressive unfreezing (unfreeze one layer every 5 epochs)");
70    let transfer_progressive = QuantumTransferLearning::new(
71        pretrained,
72        new_layers,
73        TransferStrategy::ProgressiveUnfreezing { unfreeze_rate: 5 },
74    )?;
75
76    // Step 4: Generate synthetic training data for the new task
77    println!("\n4. Generating synthetic training data...");
78    let num_samples = 50;
79    let num_features = 4;
80    let training_data = Array2::from_shape_fn((num_samples, num_features), |(i, j)| {
81        (i as f64).mul_add(0.1, j as f64 * 0.2).sin()
82    });
83    let labels = Array1::from_shape_fn(num_samples, |i| if i % 2 == 0 { 0.0 } else { 1.0 });
84
85    // Step 5: Train with fine-tuning strategy
86    println!("\n5. Training with fine-tuning strategy...");
87    let mut optimizer = Adam::new(0.01);
88
89    let result = transfer_finetune.train(
90        &training_data,
91        &labels,
92        &mut optimizer,
93        20, // epochs
94        10, // batch_size
95    )?;
96
97    println!("   Training complete!");
98    println!("   - Final loss: {:.4}", result.final_loss);
99    println!("   - Accuracy: {:.2}%", result.accuracy * 100.0);
100
101    // Step 6: Extract features using pre-trained layers
102    println!("\n6. Extracting features from pre-trained layers...");
103    let features = transfer_feature.extract_features(&training_data)?;
104    println!("   Extracted feature dimensions: {:?}", features.dim());
105
106    // Step 7: Demonstrate model zoo
107    println!("\n7. Available pre-trained models in the zoo:");
108    println!("   - Image classifier (4 qubits, MNIST subset)");
109    println!("   - Chemistry model (6 qubits, molecular energy)");
110
111    // Load chemistry model
112    let chemistry_model = QuantumModelZoo::get_chemistry_model()?;
113    println!("\n   Chemistry model info:");
114    println!("   - Task: {}", chemistry_model.task_description);
115    println!(
116        "   - MAE: {:.4}",
117        chemistry_model
118            .performance_metrics
119            .get("mae")
120            .unwrap_or(&0.0)
121    );
122    println!(
123        "   - R² score: {:.4}",
124        chemistry_model
125            .performance_metrics
126            .get("r2_score")
127            .unwrap_or(&0.0)
128    );
129
130    println!("\n=== Transfer Learning Demo Complete ===");
131
132    Ok(())
133}
134
135// Helper function to visualize layer configurations
136fn print_layer_configs(configs: &[LayerConfig]) {
137    for (i, config) in configs.iter().enumerate() {
138        println!(
139            "   Layer {}: frozen={}, lr_multiplier={:.2}",
140            i, config.frozen, config.learning_rate_multiplier
141        );
142    }
143}
transfer_learning/transfer_learning.rs

transfer_learning/
transfer_learning.rs
