crossEntropy

/* pytorch equivalent
   >>> x = torch.tensor([[0.1, 0.2], [0.3, 0.4], [0.5, 0.6]], requires_grad=True)
   >>> t = torch.tensor([1, 0, -100], dtype=torch.long)
   >>> l = torch.nn.functional.cross_entropy(x, t)
   >>> l
   tensor(0.6944)
   >>> l.backward()
   >>> x.grad
   tensor([[ 0.2375, -0.2375],
           [-0.2625,  0.2625],
           [ 0.0000,  0.0000]])
 */
import std.stdio;
import std.typecons;
import mir.ndslice;
import grain.autograd;
import numir;
static import grain.config;
grain.config.backprop = true;

auto hx = [[0.1f, 0.2f], [0.3f, 0.4f], [0.5f, 0.6f]].variable(true);
auto ht = [1, 0, -100].variable;
auto hl = crossEntropy(hx, ht);
hl.backward();
assert(approxEqual(hx.gradSliced, [[0.2375, -0.2375], [-0.2625, 0.2625],
        [0.0000, 0.0000]].nparray));

version (grain_cuda) {
    auto dx = hx.to!DeviceStorage;
    dx.grad.zero_();
    auto dt = ht.to!DeviceStorage;
    auto dl = crossEntropy(dx, dt);
    assert(approxEqual(hl.sliced, dl.to!HostStorage.sliced));
    dl.backward();
    assert(approxEqual(dx.to!HostStorage.gradSliced, hx.gradSliced));
}