def test_no_pos(self): pos_scores = torch.empty((0, ), requires_grad=True) neg_scores = torch.empty((0, 0), requires_grad=True) loss_fn = LogisticLossFunction() loss = loss_fn(pos_scores, neg_scores) self.assertTensorEqual(loss, torch.zeros(())) loss.backward()
def test_forward_bad(self): pos_scores = torch.full((3, ), -1e9, requires_grad=True) neg_scores = torch.full((3, 5), +1e9, requires_grad=True) loss_fn = LogisticLossFunction() loss = loss_fn(pos_scores, neg_scores) self.assertTensorEqual(loss, torch.tensor(6e9)) loss.backward()
def test_no_neg(self): pos_scores = torch.zeros((3, ), requires_grad=True) neg_scores = torch.empty((3, 0), requires_grad=True) loss_fn = LogisticLossFunction() loss = loss_fn(pos_scores, neg_scores, None) self.assertTensorEqual(loss, torch.tensor(2.0794)) loss.backward()
def test_forward(self): pos_scores = torch.tensor([0.8181, 0.5700, 0.3506], requires_grad=True) neg_scores = torch.tensor([ [0.4437, 0.6573, 0.9986, 0.2548, 0.0998], [0.6175, 0.4061, 0.4582, 0.5382, 0.3126], [0.9869, 0.2028, 0.1667, 0.0044, 0.9934], ], requires_grad=True) loss_fn = LogisticLossFunction() loss = loss_fn(pos_scores, neg_scores) self.assertTensorEqual(loss, torch.tensor(4.2589)) loss.backward() self.assertTrue((pos_scores.grad != 0).any()) self.assertTrue((neg_scores.grad != 0).any())