hopfield network implementation in C++
mkdir build && cd build
cmake ..
makeThis repository uses the Metal framework to offload jobs to GPU. This works only on Apple machines and it is not supported by other systems. If you are running under Linux, please compile using
make all_nometal CXX=g++it will ceate an executable called hopfield_nometal.out
Alternatively you can use cmake which automatically handles libraries
mkdir build && cd build
cmake ..
make./hopfieldOn Apple devices be sure to run from the home folder (shaders are required to be in the correct path at runtime)
// Create a dataset with memories
std::vector<Memory> training_dataset;
training_dataset.emplace_back(Memory(6, 1, 1, 0, 0, 1, 1));
training_dataset.emplace_back(Memory(6, 1, 1, 1, 1, 1, 1));
//Create a model
auto model = Model(training_dataset.size(), training_dataset[0].size());
model.load_memories(training_dataset);
// Build a corrupted memory
spin corrupted_data[] = {1, 1, 1, 1, 1, 0};
Memory corrupted_memory;
std::memmove(corrupted_memory.fData, corrupted_data, corrupted_memory.size_of());
// Train the model. Use can use kNull, kCPU, kGPU, kMultiThread, kOMP
// If the chosen device is not available another one is automatically picked up
model.train(kCPU);
model.reconstruct(corrupted_memory);
cout << "Reconstructed: " << endl;
for (uint8_t i = 0; i < corrupted_memory.size(); i++) cout << corrupted_memory.fData[i] << " ";
cout << model;More examples are in the tests folder.
Example with a corrupted image of an "8" hanwritten.
After it is reconstructed the black band disappears.
The Github Action RunMacOS sometime fails due to lack of GPU access.
See LICENSE