[metric] CRPS metric

metrics/crps/README.md

@@ -0,0 +1,34 @@
+---
+title: CRPS
+emoji: 🤗 
+colorFrom: blue
+colorTo: red
+sdk: gradio
+sdk_version: 3.19.1
+app_file: app.py
+pinned: false
+tags:
+- evaluate
+- metric
+description: >-
+  Continuous Ranked Probability Score (CRPS) is a metric that measures the accuracy of probabilistic predictions.
+---
+
+# Metric Card for CRPS
+
+## Metric description
+
+Continuous Ranked Probability Score (CRPS) is a metric that measures the accuracy of probabilistic predictions. It is commonly used in weather forecasting to measure the accuracy of predicted weather probabilities. For a random variable $X$ and a cumulative distribution function (CDF) $F$ of $X$, the CRPS is defined for a ground-truth observation $x$ and an empirical estimate of $F$ from predicted samples as:
+
+$$
+CRPS(F, x) = \int_{\inf}^{\inf} (F(z) - \mathbb{1}_{z \geq x})^2 dz,
+$$
+
+where $\mathbb{1}_{z \geq x}$ is the indicator function being identity if argument is true or zero otherwise. The CRPS is  expressed in the same unit as the observed variable and generalizes the MAE metric to probabilistic predictions. The lower the CRPS, the better the predictions.
+
+## How to use
+
+```python
+>>> crps_metric = evaluate.load("crps")
+```
+

metrics/crps/app.py

@@ -0,0 +1,6 @@
+import evaluate
+from evaluate.utils import launch_gradio_widget
+
+
+module = evaluate.load("crps")
+launch_gradio_widget(module)

metrics/crps/crps.py

@@ -0,0 +1,133 @@
+# Copyright 2022 The HuggingFace Datasets Authors and the current dataset script contributor.
+#
+# 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.
+"""CRPS - Continuous Ranked Probability Score Metric"""
+
+import datasets
+import numpy as np
+
+import evaluate
+
+
+_CITATION = """\
+@article{doi:10.1198/016214506000001437,
+    author = {Tilmann Gneiting and Adrian E Raftery},
+    title = {Strictly Proper Scoring Rules, Prediction, and Estimation},
+    journal = {Journal of the American Statistical Association},
+    volume = {102},
+    number = {477},
+    pages = {359--378},
+    year = {2007},
+    publisher = {Taylor & Francis},
+    doi = {10.1198/016214506000001437},
+    URL = {https://doi.org/10.1198/016214506000001437},
+    eprint = {https://doi.org/10.1198/016214506000001437}
+}
+"""
+
+_DESCRIPTION = """\
+Continuous Ranked Probability Score (CRPS) is the generalization of mean absolute error to the case of probabilistic forecasts used to assess the respective accuracy of probabilistic forecasting methods.
+"""
+
+_KWARGS_DESCRIPTION = """
+Args:
+    predictions: array-like of shape (n_samples,) or (n_samples, n_timesteps, n_outputs)
+        n_samples from estimated target distribution.
+    references: array-like of shape (1,) or (n_timesteps, n_outputs)
+        Empirical (correct) target values from ground truth distribution.
+    quantiles: list of floats, default=[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
+        List of quantiles in the unit interval to compute CRPS over.
+    sum: bool, default=False
+        Defines whether to sum over sum_axis dimension.
+    sum_axis: int, default=-1
+        Defines axis to sum over in case of n_outputs > 1.
+    multioutput: {"raw_values", "uniform_average"}
+        Defines aggregating across the n_outputs dimension.
+        "raw_values" returns full set of scores in case of multioutput input.
+        "uniform_average" returns the average score across all outputs.
+Returns:
+    crps: float
+        Continuous Ranked Probability Score.
+
+Examples:
+
+    >>> crps_metric = evaluate.load("crps")
+    >>> predictions = np.array([[0.1, 0.2, 0.3, 0.4, 0.5]])
+    >>> references = np.array([0.3])
+    >>> results = crps_metric.compute(predictions=predictions, references=references)
+    >>> print(results)
+    {'crps': 1.9999999254941967}
+
+    >>> crps_metric = evaluate.load("crps", "multilist")
+    >>> predictions = [[[0, 2, 4]], [[-1, 2, 5]], [[8, -5, 6]]]
+    >>> references = [[0.5], [-1], [7]]
+    >>> print(results)
+    {'crps': 3.0522875816993467}
+"""
+
+
+@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
+class Crps(evaluate.Metric):
+    def _info(self):
+        return evaluate.MetricInfo(
+            description=_DESCRIPTION,
+            citation=_CITATION,
+            inputs_description=_KWARGS_DESCRIPTION,
+            features=datasets.Features(self._get_feature_types()),
+            reference_urls=["https://www.lokad.com/continuous-ranked-probability-score/"],
+        )
+
+    def _get_feature_types(self):
+        if self.config_name == "multilist":
+            return {
+                "predictions": datasets.Sequence(datasets.Sequence(datasets.Value("float"))),
+                "references": datasets.Sequence(datasets.Value("float")),
+            }
+        else:
+            return {
+                "predictions": datasets.Sequence(datasets.Value("float")),
+                "references": datasets.Value("float"),
+            }
+
+    @staticmethod
+    def quantile_loss(target: np.ndarray, forecast: np.ndarray, q: float) -> float:
+        return 2.0 * np.sum(np.abs((target - forecast) * ((target <= forecast) - q)))
+
+    def _compute(
+        self,
+        predictions,
+        references,
+        quantiles=[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9],
+        sum=False,
+        sum_axis=-1,
+        multioutput="uniform_average",
+    ):
+        # if the number of dims of predictions > 2 then sum over sum_axis dimension if sum is True
+        if sum and len(predictions.shape) > 1:
+            predictions = np.sum(predictions, axis=sum_axis)
+            references = np.sum(references, axis=sum_axis)
+
+        abs_target_sum = np.sum(np.abs(references))
+        weighted_quantile_loss = []
+        for q in quantiles:
+            forecast_quantile = np.quantile(predictions, q, axis=0)
+            weighted_quantile_loss.append(self.quantile_loss(references, forecast_quantile, q) / abs_target_sum)
+
+        if multioutput == "raw_values":
+            return {"crps": weighted_quantile_loss}
+        elif multioutput == "uniform_average":
+            return {"crps": np.average(weighted_quantile_loss)}
+        else:
+            raise ValueError(
+                "The multioutput parameter should be one of the following: " + "'raw_values', 'uniform_average'"
+            )

metrics/crps/requirements.txt

@@ -0,0 +1,2 @@
+git+https://github.com/huggingface/evaluate@{COMMIT_PLACEHOLDER}
+numpy