Dice Metric args are unclear

[sh3rlock14]

Hi there!
I was implementing a 3D volume segmentation model and hence I wanted to use torchmetrics Dice Score...
Alas, when I tried to use Dice I discovered that by setting some args, such as average = None, even if the docs let intend that is allowed, it is not!
Indeed: at row 140 there's:
average: Optional[Literal["micro", "macro", "weighted", "none"]] = "micro",,

but at row 166 we have:

if average not in ["micro", "macro", "samples"]:
    raise ValueError(f"The `reduce` {average} is not valid.")

I was also wondering if your Dice support the preds and the g_truth to be in "one hot format" (not one hot encoding: I am in a multi-class, multi-label scenario, so I can have more than one "1" along C dim).

Thanks in advance for your help 🙏🏻 !
Mattia

[blazejdolicki]

I also stumbled upon the same issue, I think this is a bug in the library, see my comment: #1425 (comment)

[Borda]

@sh3rlock14 @blazejdolicki — this has been resolved by a complete API overhaul. The old Dice metric with confusing average/mdmc_average params has been replaced.

Current API (v1.9.0) — use torchmetrics.segmentation.DiceScore:

from torchmetrics.segmentation import DiceScore

# For 3D volume segmentation (N, C, D, H, W) one-hot format:
dice = DiceScore(
    num_classes=5,
    average="macro",            # "micro", "macro", "weighted", "none"
    input_format="one-hot",     # expects (N, C, ...) one-hot encoded
    include_background=False,   # skip background class
    per_class=False,            # True to get per-class scores
)

score = dice(preds, target)

For multi-label / overlapping classes (multiple "1"s along channel dim):
input_format="one-hot" handles this correctly — each channel is an independent binary mask:

dice = DiceScore(
    num_classes=C,
    average="none",
    input_format="one-hot",
    include_background=False,
)
per_class_dice = dice(preds.long(), target.long())

The old average="none" vs average="samples" confusion is gone — standard naming now, consistent with scikit-learn.

Docs: DiceScore