def test_loss(self):
metric = Loss()
self.assertEqual(metric.result(), 0)
metric.update(torch.tensor(1.), self.batch_size)
self.assertGreaterEqual(metric.result(), 0)
metric.reset()
self.assertEqual(metric.result(), 0)
def test_loss_multi_task(self):
metric = Loss()
self.assertEqual(metric.result(), {})
metric.update(torch.tensor(1.0), 1, 0)
metric.update(torch.tensor(2.0), 1, 1)
out = metric.result()
for k, v in out.items():
self.assertIn(k, [0, 1])
if k == 0:
self.assertEqual(v, 1)
else:
self.assertEqual(v, 2)
metric.reset()
self.assertEqual(metric.result(), {})
def test_standalone_forgetting(self):
uut = Loss()
# Initial loss should be 0
self.assertEqual(0.0, uut.result())
loss = torch.as_tensor([0.0, 0.1, 0.2, 0.3, 0.4]) # Avg = 0.2
uut.update(loss, 5)
self.assertAlmostEqual(0.2, uut.result())
loss = torch.as_tensor([0.1, 0.2, 0.3, 0.4, 0.5]) # Avg = 0.3
uut.update(loss, 5)
self.assertAlmostEqual(0.25, uut.result())
# After-reset loss should be 0
uut.reset()
self.assertEqual(0.0, uut.result())
# Check if handles 0 loss
loss = torch.as_tensor([0.0, 0.0, 0.0])
uut.update(loss, 3)
self.assertEqual(0.0, uut.result())
# Check handles losses with different reductions
uut.reset()
loss = torch.rand((5, 20))
expected_mean = loss.mean().item()
uut.update(loss, loss.shape[0])
self.assertAlmostEqual(expected_mean, uut.result())
# Check that the last call to result didn't change the value
self.assertAlmostEqual(expected_mean, uut.result())