Dice score as metric

CHANGELOG.md

@@ -26,6 +26,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Added support for nested metric collections ([#1003](https://github.com/PyTorchLightning/metrics/pull/1003))
 
 
+- Added `Dice` to classification package ([#1021](https://github.com/PyTorchLightning/metrics/pull/1021))
+
+
 ### Changed
 
 -

docs/source/classification/dice.rst

@@ -0,0 +1,33 @@
+.. customcarditem::
+   :header: Dice
+   :image: https://pl-flash-data.s3.amazonaws.com/assets/thumbnails/tabular_classification.svg
+   :tags: Classification
+
+.. include:: ../links.rst
+
+####
+Dice
+####
+
+Module Interface
+________________
+
+.. autoclass:: torchmetrics.Dice
+    :noindex:
+
+Functional Interface
+____________________
+
+.. autofunction:: torchmetrics.functional.dice
+    :noindex:
+
+
+##########
+Dice Score
+##########
+
+Functional Interface (was deprecated in v0.9)
+_____________________________________________
+
+.. autofunction:: torchmetrics.functional.dice_score
+    :noindex:

docs/source/links.rst

@@ -22,6 +22,7 @@
 .. _Matthews correlation coefficient: https://en.wikipedia.org/wiki/Matthews_correlation_coefficient
 .. _Precision: https://en.wikipedia.org/wiki/Precision_and_recall
 .. _Recall: https://en.wikipedia.org/wiki/Precision_and_recall
+.. _Dice: https://en.wikipedia.org/wiki/Sørensen–Dice_coefficient
 .. _Specificity: https://en.wikipedia.org/wiki/Sensitivity_and_specificity
 .. _Type I and Type II errors: https://en.wikipedia.org/wiki/Type_I_and_type_II_errors
 .. _confusion matrix: https://en.wikipedia.org/wiki/Confusion_matrix#Table_of_confusion

tests/classification/test_dice.py

@@ -11,10 +11,58 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from functools import partial
+from typing import Optional
+
 import pytest
-from torch import tensor
+from scipy.spatial.distance import dice as _sc_dice
+from torch import Tensor, tensor
+
+from tests.classification.inputs import _input_binary, _input_binary_logits, _input_binary_prob
+from tests.classification.inputs import _input_multiclass as _input_mcls
+from tests.classification.inputs import _input_multiclass_logits as _input_mcls_logits
+from tests.classification.inputs import _input_multiclass_prob as _input_mcls_prob
+from tests.classification.inputs import _input_multiclass_with_missing_class as _input_miss_class
+from tests.classification.inputs import _input_multilabel as _input_mlb
+from tests.classification.inputs import _input_multilabel_logits as _input_mlb_logits
+from tests.classification.inputs import _input_multilabel_multidim as _input_mlmd
+from tests.classification.inputs import _input_multilabel_multidim_prob as _input_mlmd_prob
+from tests.classification.inputs import _input_multilabel_prob as _input_mlb_prob
+from tests.helpers import seed_all
+from tests.helpers.testers import MetricTester
+from torchmetrics import Dice
+from torchmetrics.functional import dice, dice_score
+from torchmetrics.functional.classification.stat_scores import _del_column
+from torchmetrics.utilities.checks import _input_format_classification
+from torchmetrics.utilities.enums import DataType
+
+seed_all(42)
+
+
+def _sk_dice(
+    preds: Tensor,
+    target: Tensor,
+    ignore_index: Optional[int] = None,
+) -> float:
+    """Compute dice score from prediction and target. Used scipy implementation of main dice logic.
 
-from torchmetrics.functional import dice_score
+    Args:
+        preds: prediction tensor
+        target: target tensor
+        ignore_index:
+            Integer specifying a target class to ignore. Recommend set to index of background class.
+    Return:
+        Float dice score
+    """
+    sk_preds, sk_target, mode = _input_format_classification(preds, target)
+
+    if ignore_index is not None and mode != DataType.BINARY:
+        sk_preds = _del_column(sk_preds, ignore_index)
+        sk_target = _del_column(sk_target, ignore_index)
+
+    sk_preds, sk_target = sk_preds.numpy(), sk_target.numpy()
+
+    return 1 - _sc_dice(sk_preds.reshape(-1), sk_target.reshape(-1))
 
 
 @pytest.mark.parametrize(
@@ -29,3 +77,89 @@
 def test_dice_score(pred, target, expected):
     score = dice_score(tensor(pred), tensor(target))
     assert score == expected
+
+
+@pytest.mark.parametrize(
+    ["pred", "target", "expected"],
+    [
+        ([[0, 0], [1, 1]], [[0, 0], [1, 1]], 1.0),
+        ([[1, 1], [0, 0]], [[0, 0], [1, 1]], 0.0),
+        ([[1, 1], [1, 1]], [[1, 1], [0, 0]], 2 / 3),
+        ([[1, 1], [0, 0]], [[1, 1], [0, 0]], 1.0),
+    ],
+)
+def test_dice(pred, target, expected):
+    score = dice(tensor(pred), tensor(target), ignore_index=0)
+    assert score == expected
+
+
+@pytest.mark.parametrize(
+    "preds, target",
+    [
+        (_input_binary.preds, _input_binary.target),
+        (_input_binary_logits.preds, _input_binary_logits.target),
+        (_input_binary_prob.preds, _input_binary_prob.target),
+    ],
+)
+@pytest.mark.parametrize("ignore_index", [None])
+class TestDiceBinary(MetricTester):
+    @pytest.mark.parametrize("ddp", [False])
+    @pytest.mark.parametrize("dist_sync_on_step", [False])
+    def test_dice_class(self, ddp, dist_sync_on_step, preds, target, ignore_index):
+        self.run_class_metric_test(
+            ddp=ddp,
+            preds=preds,
+            target=target,
+            metric_class=Dice,
+            sk_metric=partial(_sk_dice, ignore_index=ignore_index),
+            dist_sync_on_step=dist_sync_on_step,
+            metric_args={"ignore_index": ignore_index},
+        )
+
+    def test_dice_fn(self, preds, target, ignore_index):
+        self.run_functional_metric_test(
+            preds,
+            target,
+            metric_functional=dice,
+            sk_metric=partial(_sk_dice, ignore_index=ignore_index),
+            metric_args={"ignore_index": ignore_index},
+        )
+
+
+@pytest.mark.parametrize(
+    "preds, target",
+    [
+        (_input_mcls.preds, _input_mcls.target),
+        (_input_mcls_logits.preds, _input_mcls_logits.target),
+        (_input_mcls_prob.preds, _input_mcls_prob.target),
+        (_input_miss_class.preds, _input_miss_class.target),
+        (_input_mlb.preds, _input_mlb.target),
+        (_input_mlb_logits.preds, _input_mlb_logits.target),
+        (_input_mlmd.preds, _input_mlmd.target),
+        (_input_mlmd_prob.preds, _input_mlmd_prob.target),
+        (_input_mlb_prob.preds, _input_mlb_prob.target),
+    ],
+)
+@pytest.mark.parametrize("ignore_index", [None, 0])
+class TestDiceMulti(MetricTester):
+    @pytest.mark.parametrize("ddp", [False])
+    @pytest.mark.parametrize("dist_sync_on_step", [False])
+    def test_dice_class(self, ddp, dist_sync_on_step, preds, target, ignore_index):
+        self.run_class_metric_test(
+            ddp=ddp,
+            preds=preds,
+            target=target,
+            metric_class=Dice,
+            sk_metric=partial(_sk_dice, ignore_index=ignore_index),
+            dist_sync_on_step=dist_sync_on_step,
+            metric_args={"ignore_index": ignore_index},
+        )
+
+    def test_dice_fn(self, preds, target, ignore_index):
+        self.run_functional_metric_test(
+            preds,
+            target,
+            metric_functional=dice,
+            sk_metric=partial(_sk_dice, ignore_index=ignore_index),
+            metric_args={"ignore_index": ignore_index},
+        )

torchmetrics/__init__.py

@@ -33,6 +33,7 @@
     CohenKappa,
     ConfusionMatrix,
     CoverageError,
+    Dice,
     F1Score,
     FBetaScore,
     HammingDistance,
@@ -126,6 +127,7 @@
     "ConfusionMatrix",
     "CosineSimilarity",
     "CoverageError",
+    "Dice",
     "TweedieDevianceScore",
     "ErrorRelativeGlobalDimensionlessSynthesis",
     "ExplainedVariance",

torchmetrics/classification/__init__.py

@@ -21,6 +21,7 @@
 from torchmetrics.classification.calibration_error import CalibrationError  # noqa: F401
 from torchmetrics.classification.cohen_kappa import CohenKappa  # noqa: F401
 from torchmetrics.classification.confusion_matrix import ConfusionMatrix  # noqa: F401
+from torchmetrics.classification.dice import Dice  # noqa: F401
 from torchmetrics.classification.f_beta import F1Score, FBetaScore  # noqa: F401
 from torchmetrics.classification.hamming import HammingDistance  # noqa: F401
 from torchmetrics.classification.hinge import HingeLoss  # noqa: F401

torchmetrics/classification/dice.py

@@ -0,0 +1,155 @@
+# Copyright The PyTorch Lightning team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, Optional
+
+from torch import Tensor
+
+from torchmetrics.classification.stat_scores import StatScores
+from torchmetrics.functional.classification.dice import _dice_compute
+
+
+class Dice(StatScores):
+    r"""Computes `Dice`_:
+
+    .. math:: \text{Dice} = \frac{\text{2 * TP}}{\text{2 * TP} + \text{FP} + \text{FN}}
+
+    Where :math:`\text{TP}` and :math:`\text{FP}` represent the number of true positives and
+    false positives respecitively.
+
+    It is recommend set `ignore_index` to index of background class.
+
+    The reduction method (how the precision scores are aggregated) is controlled by the
+    ``average`` parameter, and additionally by the ``mdmc_average`` parameter in the
+    multi-dimensional multi-class case. Accepts all inputs listed in :ref:`pages/classification:input types`.
+
+    Args:
+        num_classes:
+            Number of classes. Necessary for ``'macro'``, ``'weighted'`` and ``None`` average methods.
+        threshold:
+            Threshold for transforming probability or logit predictions to binary (0,1) predictions, in the case
+            of binary or multi-label inputs. Default value of 0.5 corresponds to input being probabilities.
+        zero_division:
+            The value to use for the score if denominator equals zero.
+        average:
+            Defines the reduction that is applied. Should be one of the following:
+
+            - ``'micro'`` [default]: Calculate the metric globally, across all samples and classes.
+            - ``'macro'``: Calculate the metric for each class separately, and average the
+              metrics across classes (with equal weights for each class).
+            - ``'weighted'``: Calculate the metric for each class separately, and average the
+              metrics across classes, weighting each class by its support (``tp + fn``).
+            - ``'none'`` or ``None``: Calculate the metric for each class separately, and return
+              the metric for every class.
+            - ``'samples'``: Calculate the metric for each sample, and average the metrics
+              across samples (with equal weights for each sample).
+
+            .. note:: What is considered a sample in the multi-dimensional multi-class case
+                depends on the value of ``mdmc_average``.
+
+        mdmc_average:
+            Defines how averaging is done for multi-dimensional multi-class inputs (on top of the
+            ``average`` parameter). Should be one of the following:
+
+            - ``None`` [default]: Should be left unchanged if your data is not multi-dimensional
+              multi-class.
+
+            - ``'samplewise'``: In this case, the statistics are computed separately for each
+              sample on the ``N`` axis, and then averaged over samples.
+              The computation for each sample is done by treating the flattened extra axes ``...``
+              (see :ref:`pages/classification:input types`) as the ``N`` dimension within the sample,
+              and computing the metric for the sample based on that.
+
+            - ``'global'``: In this case the ``N`` and ``...`` dimensions of the inputs
+              (see :ref:`pages/classification:input types`) are flattened into a new ``N_X`` sample axis, i.e.
+              the inputs are treated as if they were ``(N_X, C)``.
+              From here on the ``average`` parameter applies as usual.
+
+        ignore_index:
+            Integer specifying a target class to ignore. If given, this class index does not contribute
+            to the returned score, regardless of reduction method. If an index is ignored, and ``average=None``
+            or ``'none'``, the score for the ignored class will be returned as ``nan``.
+
+        top_k:
+            Number of the highest probability or logit score predictions considered finding the correct label,
+            relevant only for (multi-dimensional) multi-class inputs. The
+            default value (``None``) will be interpreted as 1 for these inputs.
+            Should be left at default (``None``) for all other types of inputs.
+
+        multiclass:
+            Used only in certain special cases, where you want to treat inputs as a different type
+            than what they appear to be. See the parameter's
+            :ref:`documentation section <pages/classification:using the multiclass parameter>`
+            for a more detailed explanation and examples.
+
+        kwargs: Additional keyword arguments, see :ref:`Metric kwargs` for more info.
+
+    Raises:
+        ValueError:
+            If ``average`` is none of ``"micro"``, ``"macro"``, ``"weighted"``, ``"samples"``, ``"none"``, ``None``.
+
+    Example:
+        >>> import torch
+        >>> from torchmetrics import Dice
+        >>> preds  = torch.tensor([2, 0, 2, 1])
+        >>> target = torch.tensor([1, 1, 2, 0])
+        >>> dice = Dice(average='micro')
+        >>> dice(preds, target)
+        tensor(0.2500)
+
+    """
+    is_differentiable = False
+    higher_is_better = True
+
+    def __init__(
+        self,
+        zero_division: int = 0,
+        num_classes: Optional[int] = None,
+        threshold: float = 0.5,
+        average: str = "micro",
+        mdmc_average: Optional[str] = "global",
+        ignore_index: Optional[int] = None,
+        top_k: Optional[int] = None,
+        multiclass: Optional[bool] = None,
+        **kwargs: Dict[str, Any],
+    ) -> None:
+        allowed_average = ("micro", "macro", "weighted", "samples", "none", None)
+        if average not in allowed_average:
+            raise ValueError(f"The `average` has to be one of {allowed_average}, got {average}.")
+
+        super().__init__(
+            reduce="macro" if average in ("weighted", "none", None) else average,
+            mdmc_reduce=mdmc_average,
+            threshold=threshold,
+            top_k=top_k,
+            num_classes=num_classes,
+            multiclass=multiclass,
+            ignore_index=ignore_index,
+            **kwargs,
+        )
+
+        self.average = average
+        self.zero_division = zero_division
+
+    def compute(self) -> Tensor:
+        """Computes the dice score based on inputs passed in to ``update`` previously.
+
+        Return:
+            The shape of the returned tensor depends on the ``average`` parameter:
+
+            - If ``average in ['micro', 'macro', 'weighted', 'samples']``, a one-element tensor will be returned
+            - If ``average in ['none', None]``, the shape will be ``(C,)``, where ``C`` stands  for the number
+              of classes
+        """
+        tp, fp, _, fn = self._get_final_stats()
+        return _dice_compute(tp, fp, fn, self.average, self.mdmc_reduce, self.zero_division)