def test_class_reduce():
num = torch.randint(1, 10, (100, )).float()
denom = torch.randint(10, 20, (100, )).float()
weights = torch.randint(1, 100, (100, )).float()
assert torch.allclose(class_reduce(num, denom, weights, 'micro'),
torch.sum(num) / torch.sum(denom))
assert torch.allclose(class_reduce(num, denom, weights, 'macro'),
torch.mean(num / denom))
assert torch.allclose(
class_reduce(num, denom, weights, 'weighted'),
torch.sum(num / denom * (weights / torch.sum(weights))))
assert torch.allclose(class_reduce(num, denom, weights, 'none'),
num / denom)
def _fbeta_compute(true_positives: torch.Tensor,
predicted_positives: torch.Tensor,
actual_positives: torch.Tensor,
beta: float = 1.0,
average: str = "micro") -> torch.Tensor:
if average == "micro":
precision = true_positives.sum().float() / predicted_positives.sum()
recall = true_positives.sum().float() / actual_positives.sum()
else:
precision = true_positives.float() / predicted_positives
recall = true_positives.float() / actual_positives
num = (1 + beta**2) * precision * recall
denom = beta**2 * precision + recall
return class_reduce(num,
denom,
weights=actual_positives,
class_reduction=average)