def test_accuracy():
acc_calculator = Accuracy(average=False, topk=(2, ))
# start with multiclass classification
pred = torch.FloatTensor([[0.4, 0.5, 0.1], [0.1, 0.8, 0.1]])
target = torch.LongTensor([[2], [1]])
assert pred.size() == torch.Size([2, 3])
# check for length of the data
for i in range(2):
acc_calculator.update(output=pred[i, :].unsqueeze(0),
target=target[i, :])
assert acc_calculator.value.shape == (2, 1)
assert np.allclose(acc_calculator.value, np.array([[0], [1]]))
acc_calculator = Accuracy(average=True, topk=(2, ))
# check for length of the data
for i in range(2):
acc_calculator.update(output=pred[i, :].unsqueeze(0),
target=target[i, :])
assert acc_calculator.value == 0.5
def test_that_accuracy_metric_calculates_multiclass_correctly(y_ypred):
predicted, true = y_ypred
predicted = predicted.softmax(-1) + torch.randn_like(predicted) / 100
# test adding a batch:
accuracy_calculator = Accuracy(average=True, topk=(1, ))
accuracy_calculator.update(predicted, true)
assert np.allclose(accuracy_calculator.value,
(true == predicted.argmax(-1)).float().mean().item())
# test adding all elements in a batch separately
new_accuracy_calculator = Accuracy(average=True, topk=(1, ))
for x, y in zip(predicted, true):
x, y = x.unsqueeze(0), y.unsqueeze(0)
new_accuracy_calculator.update(x, y)
assert np.allclose(new_accuracy_calculator.value,
accuracy_calculator.value)
def test_that_accuracy_metric_calculates_top_k_correctly(y_ypred):
predicted, true = y_ypred
predicted = predicted.softmax(-1) + torch.randn_like(predicted) / 100
last_accuracy = 0
for k in range(1, predicted.size(1) + 1):
accuracy_calculator = Accuracy(average=True, topk=(k, ))
accuracy_calculator.update(predicted, true)
assert 1 >= accuracy_calculator.value >= last_accuracy
last_accuracy = accuracy_calculator.value
accuracy_calculator = Accuracy(average=True,
topk=tuple(range(1,
predicted.size(1) + 1)))
accuracy_calculator.update(predicted, true)
assert all(val_1 >= val_0 for val_0, val_1 in zip(
accuracy_calculator.value[::2], accuracy_calculator.value[1::2]))
def test_that_accuracy_raises_errors_when_shapes_dont_match():
predicted = torch.randn(5, 3)
true = torch.tensor([0, 0, 0])
accuracy_calculator = Accuracy(average=True, topk=(1, ))
with pytest.raises(ValueError):
accuracy_calculator.update(predicted, true)
def test_that_accuracy_string_repr_doesnt_throw_errors(y_ypred):
predicted, true = y_ypred
predicted = predicted.softmax(-1) + torch.randn_like(predicted) / 100
accuracy = Accuracy(average=True, topk=(1, 2))
accuracy.update(predicted, true)
assert "±" in str(accuracy)
