def test_cm(self):
metric = ConfusionMatrix()
cm = metric.result()
self.assertTrue((cm == 0).all().item())
metric.update(self.y, self.out)
cm = metric.result()
self.assertTrue((cm >= 0).all().item())
metric.reset()
cm = metric.result()
self.assertTrue((cm == 0).all().item())
def test_standalone_cm_error_handling(self):
uut = ConfusionMatrix(num_classes=10)
# Initial confusion matrix should be a 10x10 matrix filled with zeros
self.assertTrue(
torch.equal(torch.zeros((10, 10), dtype=torch.long), uut.result()))
truth = torch.as_tensor([0, 5, 2, 1, 0])
predicted = torch.as_tensor([2, 3, -1, 5, 0])
with self.assertRaises(ValueError):
uut.update(truth, predicted)
truth = torch.as_tensor([0, 5, 2, 1, -5])
predicted = torch.as_tensor([2, 3, 1, 5, 0])
with self.assertRaises(ValueError):
uut.update(truth, predicted)
def test_standalone_cm_fixed_size(self):
uut = ConfusionMatrix(num_classes=10)
# Initial confusion matrix should be a 10x10 matrix filled with zeros
self.assertTrue(
torch.equal(torch.zeros((10, 10), dtype=torch.long), uut.result()))
truth = torch.as_tensor([0, 5, 2, 1, 0])
predicted = torch.as_tensor([2, 3, 2, 5, 0])
uut.update(truth, predicted)
expected = torch.zeros((10, 10), dtype=torch.long)
expected[0][2] += 1
expected[5][3] += 1
expected[2][2] += 1
expected[1][5] += 1
expected[0][0] += 1
self.assertTrue(torch.equal(expected, uut.result()))
# After-reset matrix should be a 10x10 matrix filled with zeros
uut.reset()
self.assertTrue(
torch.equal(torch.zeros((10, 10), dtype=torch.long), uut.result()))
# Should throw exception when len(truth) != len(predicted)
with self.assertRaises(ValueError):
truth = torch.as_tensor([0, 0, 1, 0])
predicted = torch.as_tensor([1, 1, 1])
uut.update(truth, predicted)
# Check that matrix didn't change after error
self.assertTrue(
torch.equal(torch.zeros((10, 10), dtype=torch.long), uut.result()))
# Test logits / one-hot support
uut.reset()
# Test one-hot (truth)
truth = torch.as_tensor([[0, 1], [1, 0], [1, 0], [1, 0]])
predicted = torch.as_tensor([1, 1, 1, 1])
uut.update(truth, predicted)
expected = torch.zeros((10, 10), dtype=torch.long)
expected[1][1] += 1
expected[0][1] += 1
expected[0][1] += 1
expected[0][1] += 1
self.assertTrue(torch.equal(expected, uut.result()))
# Test logits (predictions)
truth = torch.as_tensor([1, 1, 0, 0])
predicted = torch.as_tensor([[0.73, 0.1], [0.22, 0.33], [0.99, 0.01],
[0.12, 0.11]])
uut.update(truth, predicted)
expected[1][0] += 1
expected[1][1] += 1
expected[0][0] += 1
expected[0][0] += 1
self.assertTrue(torch.equal(expected, uut.result()))
# Test one-hot (truth) + logits (predictions)
truth = torch.as_tensor([[1, 0], [1, 0], [0, 1], [0, 1]])
predicted = torch.as_tensor([[0.73, 0.1], [0.22, 0.33], [0.99, 0.01],
[0.12, 0.11]])
uut.update(truth, predicted)
expected[0][0] += 1
expected[0][1] += 1
expected[1][0] += 1
expected[1][0] += 1
self.assertTrue(torch.equal(expected, uut.result()))