def test_that_precision_metric_calculates_multiclass_correctly(y_ypred):
predicted, true = y_ypred
predicted = predicted.softmax(-1) + torch.randn_like(predicted) / 100
# test adding a batch:
precision = Precision(num_classes=predicted.size(1), average=True)
precision.update(predicted, true)
true_positive = ((torch.zeros_like(predicted).scatter(
1,
predicted.argmax(-1).unsqueeze(-1), 1) > 0)
& (torch.zeros_like(predicted).scatter(
1, true.unsqueeze(-1), 1) > 0)).sum(dim=0)
all_positive = (torch.zeros_like(predicted).scatter(
1, true.unsqueeze(-1), 1) > 0).sum(dim=0)
manual_precision = true_positive.float() / all_positive.float()
manual_precision[torch.isnan(manual_precision)] = 0
# breakpoint()
assert np.allclose(precision.value, manual_precision.mean())
# test adding all elements in a batch separately
new_precision = Precision(num_classes=predicted.size(1), average=True)
for x, y in zip(predicted, true):
x, y = x.unsqueeze(0), y.unsqueeze(0)
new_precision.update(x, y)
assert np.allclose(new_precision.value, precision.value)
def test_precision():
prec_calculator = Precision(num_classes=3, average=True)
# start with multiclass classification
pred = torch.FloatTensor([[0.4, 0.5, 0.1], [0.1, 0.8, 0.1]])
target = torch.LongTensor([[2], [1]])
for i in range(2):
prec_calculator.update(output=pred[i, :].unsqueeze(0),
target=target[i, :].unsqueeze(0))
assert round(prec_calculator.value, 2) == 0.33
prec_calculator = Precision(num_classes=3, average=False)
# start with multiclass classification
pred = torch.FloatTensor([[0.4, 0.5, 0.1], [0.1, 0.8, 0.1]])
target = torch.LongTensor([[2], [1]])
for i in range(2):
prec_calculator.update(output=pred[i, :].unsqueeze(0),
target=target[i, :].unsqueeze(0))
assert np.allclose(np.array(prec_calculator.value), np.array([0, 1, 0]))
def test_that_precision_string_repr_doesnt_throw_errors(y_ypred):
predicted, true = y_ypred
predicted = predicted.softmax(-1) + torch.randn_like(predicted) / 100
precision = Precision(num_classes=predicted.size(1), average=True)
precision.update(predicted, true)
assert "±" in str(precision)