make

import grain.autograd;
import numir;

{
    auto model = MLP!(float, HostStorage)(3);
    // auto optim = AdaDelta!(typeof(model))(model);
    auto optim = make!AdaDelta(model);
    // static assert(isOptimizer!(typeof(optim)));
    model.fc1.weight.data.zero_();
    model.fc1.weight.grad = [[0.2f, 0.0f, 0.0f], [0.0f, 0.0f, 0.0f]].variable
        .data;
    optim.update();
    auto w = model.fc1.weight;
    auto d = (1.0 - optim.rho) * 0.2 * 0.2;
    auto diff = cast(float)((0.0 + optim.eps) / (d + optim.eps)) ^^ 0.5;
    auto n = (1.0 - optim.rho) * diff * diff;
    assert(approxEqual(w.sliced, [[-optim.lr * diff, -optim.lr,
            -optim.lr], [-optim.lr, -optim.lr, -optim.lr]].nparray));
    auto d_ = optim.den[".fc1.weight"].to!(typeof(w));
    auto n_ = optim.num[".fc1.weight"].to!(typeof(w));
    assert(approxEqual(d_.sliced, [[d, 0.0, 0.0], [0.0, 0.0, 0.0]].nparray));
    assert(approxEqual(n_.sliced[0, 0 .. 1], [n].nparray));
}
version (grain_cuda) {
    auto model = MLP!(float, DeviceStorage)(3);
    auto optim = AdaDelta!(typeof(model))(model);
    optim.update();
}