Implementing a Custom Expansion Strategy

ExpansionStrategy<T> controls the iteration loop: which labels to expand, in what order and when to stop. Subclass it to implement a custom policy.

Interface

// src/strategies/ExpansionStrategy.hpp
template <typename T>
class ExpansionStrategy {
public:
    virtual int execute(AlphaExpansion<T>& optimizer, EnergyModel<T>& model) const = 0;
    virtual ~ExpansionStrategy() = default;
};

execute() calls optimizer.perform_expansion_move(alpha) for whatever labels and in whatever order the strategy decides. It returns the number of full cycles completed.

Example: Single-Pass Strategy

Tries each label exactly once and stops:

// src/strategies/SinglePassStrategy.hpp
#pragma once
 
#include "strategies/ExpansionStrategy.hpp"
#include "core/AlphaExpansion.hpp"
#include "core/EnergyModel.hpp"
 
template <typename T>
class SinglePassStrategy : public ExpansionStrategy<T> {
public:
    int execute(AlphaExpansion<T>& optimizer, EnergyModel<T>& model) const override {
        for (int alpha = 0; alpha < model.num_labels(); ++alpha)
            optimizer.perform_expansion_move(alpha);
        return 1;
    }
};

Using the Custom Strategy

#include "core/EnergyModel.hpp"
#include "core/AlphaExpansion.hpp"
#include "solvers/BKSolver.hpp"
#include "strategies/SinglePassStrategy.hpp"
 
EnergyModel<int> model(100, 3);
// ... set up costs and graph ...
 
auto factory = [](int v, int e) {
    return std::make_unique<BKSolver<int>>(v, e);
};
AlphaExpansion<int> optimizer(model, factory);
 
SinglePassStrategy<int> strategy;
int cycles = strategy.execute(optimizer, model);

Tips

• Call model.evaluate_total_energy() to check progress between moves.
• Call model.get_labels() and model.set_labels() to save and restore state. GreedyStrategy uses this to roll back non-improving moves.
• Return the number of full cycles from execute(). The demo app shows this value in the convergence dialog.