ece.py

# Copyright 2020 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.
"""TODO: Add a description here."""

# https://huggingface.co/spaces/jordyvl/ece

import evaluate
import datasets
import numpy as np


# TODO: Add BibTeX citation
_CITATION = """\
@InProceedings{huggingface:module,
title = {Expected Calibration Error},
authors={Jordy Van Landeghem},
year={2022}
}
"""

# TODO: Add description of the module here
_DESCRIPTION = """\
This new module is designed to evaluate the calibration of a probabilistic classifier.
More concretely, we provide a binned empirical estimator of top-1 calibration error. [1] 
"""


# TODO: Add description of the arguments of the module here
_KWARGS_DESCRIPTION = """
Calculates how good are predictions given some references, using certain scores
Args:
    predictions: list of predictions to score. Each predictions
        should be a string with tokens separated by spaces.
    references: list of reference for each prediction. Each
        reference should be a string with tokens separated by spaces.
    y_true : array-like
        Ground truth labels.
    p_hat : array-like
        Array of confidence estimates.
    n_bins : int, default=15
        Number of bins of :math:`[\\frac{1}{n_{\\text{classes}},1]` for the confidence estimates.
    n_classes : int default=None
        Number of classes. Estimated from `y` and `y_pred` if not given.
    p : int, default=1
        Power of the calibration error, :math:`1 \\leq p \\leq \\infty`.



Returns
    Expected calibration error (ECE), float.

Examples:
    Examples should be written in doctest format, and should illustrate how
    to use the function.

    >>> my_new_module = evaluate.load("jordyvl/ece")
    >>> results = my_new_module.compute(references=[0, 1, 2], predictions=[[0.6, 0.2, 0.2], [0, 0.95, 0.05], [0.7, 0.1 ,0.2]])
    >>> print(results)
    {'ECE': 1.0}
"""

# TODO: Define external resources urls if needed
BAD_WORDS_URL = ""


# Discretization and binning
def create_bins(n_bins=10, scheme="equal-range", bin_range=None, P=None):
    assert scheme in [
        "equal-range",
        "equal-masss",
    ], f"This binning scheme {scheme} is not implemented yet"

    if bin_range is None:
        if P is None:
            bin_range = [0, 1]  # no way to know range
        else:
            bin_range = [min(P), max(P)]

    if scheme == "equal-range":
        bins = np.linspace(bin_range[0], bin_range[1], n_bins + 1)  # equal range
        # bins = np.tile(np.linspace(bin_range[0], bin_range[1], n_bins + 1), (n_classes,1))

    elif scheme == "equal-mass":
        assert P.size >= n_bins, "Fewer points than bins"

        # assume global equal mass binning; not discriminated per class
        P = P.flatten()

        # split sorted probabilities into groups of approx equal size
        groups = np.array_split(np.sort(P), n_bins)
        bin_upper_edges = list()

        # rightmost entry per equal size group
        for cur_group in range(n_bins - 1):
            bin_upper_edges += [max(groups[cur_group])]
        bin_upper_edges += [np.inf]  # always +1 for right edges
        bins = np.array(bin_upper_edges)

    return bins


def discretize_into_bins(P, bins):
    oneDbins = np.digitize(P, bins) - 1  # since bins contains extra righmost & leftmost bins

    # Fix to scipy.binned_dd_statistic:
    # Tie-breaking to the left for rightmost bin
    # Using `digitize`, values that fall on an edge are put in the right bin.
    # For the rightmost bin, we want values equal to the right
    # edge to be counted in the last bin, and not as an outlier.

    for k in range(P.shape[-1]):
        # Find the rounding precision
        dedges_min = np.diff(bins).min()
        if dedges_min == 0:
            raise ValueError("The smallest edge difference is numerically 0.")

        decimal = int(-np.log10(dedges_min)) + 6

        # Find which points are on the rightmost edge.
        on_edge = np.where(
            (P[:, k] >= bins[-1]) & (np.around(P[:, k], decimal) == np.around(bins[-1], decimal))
        )[0]
        # Shift these points one bin to the left.
        oneDbins[on_edge, k] -= 1

    return oneDbins


def manual_binned_statistic(P, y_correct, bins, statistic="mean"):
    bin_assignments = discretize_into_bins(np.expand_dims(P, 0), bins)[0]
    result = np.empty([len(bins)], float)
    result.fill(np.nan)  # cannot assume each bin will have observations

    flatcount = np.bincount(bin_assignments, None)
    a = flatcount.nonzero()

    if statistic == "mean":
        flatsum = np.bincount(bin_assignments, y_correct)
        result[a] = flatsum[a] / flatcount[a]
    return result, bins, bin_assignments + 1  # fix for what happens in discretize_into_bins


def bin_calibrated_accuracy(bins, proxy="upper-edge"):
    assert proxy in ["center", "upper-edge"], f"Unsupported proxy{proxy}"

    if proxy == "upper-edge":
        return bins[1:]

    if proxy == "center":
        return bins[:-1] + np.diff(bins) / 2


def CE_estimate(y_correct, P, bins=None, p=1, proxy="upper-edge"):
    """
    y_correct: binary (N x 1)
    P: normalized (N x 1) either max or per class

    Summary: weighted average over the accuracy/confidence difference of discrete bins of prediction probability
    """

    n_bins = len(bins) - 1
    bin_range = [min(bins), max(bins)]

    # average bin probability #55 for bin 50-60, mean per bin; or right/upper bin edges
    calibrated_acc = bin_calibrated_accuracy(bins, proxy="upper-edge")

    empirical_acc, bin_edges, bin_assignment = manual_binned_statistic(P, y_correct, bins)
    bin_numbers, weights_ece = np.unique(bin_assignment, return_counts=True)
    anindices = bin_numbers - 1  # reduce bin counts; left edge; indexes right by default

    # Expected calibration error
    if p < np.inf:  # L^p-CE
        CE = np.average(
            abs(empirical_acc[anindices] - calibrated_acc[anindices]) ** p, weights=weights_ece
        )
    elif np.isinf(p):  # max-ECE
        CE = np.max(abs(empirical_acc[anindices] - calibrated_acc[anindices]))

    return CE


def top_1_CE(Y, P, **kwargs):
    y_correct = (Y == np.argmax(P, -1)).astype(int)  # create condition y = ŷ € [K]
    p_max = np.max(P, -1)  # create p̂ as top-1 softmax probability € [0,1]
    bins = create_bins(
        n_bins=kwargs["n_bins"], bin_range=kwargs["bin_range"], scheme=kwargs["scheme"], P=p_max
    )
    return CE_estimate(y_correct, p_max, bins=bins, proxy=kwargs["proxy"])


@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class ECE(evaluate.EvaluationModule):
    """
    0. create binning scheme [discretization of f]
    1. build histogram P(f(X))
    2. build conditional density estimate P(y|f(X))
    3. average bin probabilities f_B as center/edge of bin
    4. apply L^p norm distance and weights
    """

    def _info(self):
        # TODO: Specifies the evaluate.EvaluationModuleInfo object
        return evaluate.EvaluationModuleInfo(
            module_type="metric",
            description=_DESCRIPTION,
            citation=_CITATION,
            inputs_description=_KWARGS_DESCRIPTION,
            features=datasets.Features(
                {
                    "predictions": datasets.Sequence(datasets.Value("float32")),
                    "references": datasets.Value("int64"),
                }
            ),
            # Homepage of the module for documentation
            homepage="https://huggingface.co/spaces/jordyvl/ece",
            # Additional links to the codebase or references
            codebase_urls=["http://github.com/path/to/codebase/of/new_module"],
            reference_urls=["http://path.to.reference.url/new_module"],
        )

    def init_kwargs(
        self, n_bins=10, bin_range=None, scheme="equal-range", proxy="upper-edge", p=1, **kwargs
    ):
        # super(evaluate.EvaluationModule, self).__init__(**kwargs)
        self.n_bins = n_bins
        self.bin_range = bin_range
        self.scheme = scheme
        self.proxy = proxy
        self.p = p

    def _compute(self, predictions, references, **kwargs):

        # convert to numpy arrays
        references = np.array(references, dtype=np.int64)
        predictions = np.array(predictions, dtype=np.float32)

        assert (
            predictions.shape[0] == references.shape[0]
        ), "Need to pass similar predictions and references"

        # Assert that arrays are 2D
        if len(predictions.shape) != 2:
            raise ValueError("Expected `predictions` to be a 2D vector (N x K)")

        if len(references.shape) != 1:
            # could check if wrongly passed as onehot
            if (references.shape[-1] == predictions.shape[1]) and (
                np.sum(references) == predictions.shape[0]
            ):
                references = np.argmax(references, -1)
            else:
                raise ValueError("Expected `references` to be a 1D vector (N,)")

        self.init_kwargs(**kwargs)

        """Returns the scores"""
        ECE = top_1_CE(references, predictions, **self.__dict__)
        return {
            "ECE": ECE,
        }