def test_one_hot(self, smooth_eps, K):
batch_size = 32
y = torch.randint(K, (batch_size, ))
y_np = y.numpy()
actual = losses.one_hot(y, K, smooth_eps)
y_np_one_hot = np.eye(K)[y_np]
expected = y_np_one_hot * (1 - smooth_eps) + (smooth_eps / (K - 1))
assert_allclose(actual, expected)
def test_one_hot_eps_out_of_bounds(self, smooth_eps):
with pytest.raises(AssertionError):
losses.one_hot(torch.randint(5, (2, 2)), 5, smooth_eps)
def test_one_hot_not_rank_one(self):
with pytest.raises(AssertionError):
losses.one_hot(torch.randint(5, (2, 2)), 5, 0)