def test_fscore_empty(test_mask_batch):
num_classes = 6
pred_classes = (test_mask_batch - 1) % num_classes
assert pred_classes.shape == (4, 4, 4)
tp, fp, tn, fn = func.statistics_step(pred_classes,
test_mask_batch,
num_classes=num_classes,
ignore_index=255,
reduction=False)
score = func.f1_score(tp, fp, fn, reduce=True)
assert score == 0.0
LOG.debug(score)
def test_fscore_best():
num_classes = 6
test_mask_batch = torch.randint(0, 6, size=(4, 4, 4))
pred_classes = test_mask_batch.clone()
assert pred_classes.shape == (4, 4, 4)
tp, fp, tn, fn = func.statistics_step(pred_classes,
test_mask_batch,
num_classes=num_classes,
ignore_index=255,
reduction=False)
score = func.f1_score(tp, fp, fn, reduce=False)
LOG.debug(score)
assert torch.all(score[:4] >= 0.99)
def compute(self) -> torch.Tensor:
"""Computes the F1 score over every device, using the accumulated statistics.
Same micro and macro-average considerations hold for this metric as well.
:return: tensor with empty size when reduced, or (C,) where C in the number of classes
:rtype: torch.Tensor
"""
score = func.f1_score(tp=self.tp,
fp=self.fp,
fn=self.fn,
reduce=self.is_micro)
if self.reduction == ReductionType.MACRO.value:
score = score.mean()
return score
def test_fscore_batch_macro(test_pred_batch, test_mask_batch):
pred_classes = test_pred_batch.argmax(dim=1)
assert pred_classes.shape == (4, 4, 4)
tp, fp, tn, fn = func.statistics_step(pred_classes,
test_mask_batch,
num_classes=6,
ignore_index=255,
reduction=False)
fscore = func.f1_score(tp, fp, fn, reduce=True)
y_true, y_pred = func.valid_samples(255, test_mask_batch, pred_classes)
skl_prec = f1_score(y_true.cpu().numpy(),
y_pred.cpu().numpy(),
average="macro")
LOG.debug("sklearn: %s - custom: %s", str(skl_prec), str(fscore))
# sklearn does not account for empty classes
diff = abs(skl_prec - fscore.item())
assert diff <= EPS