# Copyright (C) 2017-2022 Cleanlab Inc.
# This file is part of cleanlab.
# cleanlab is free software: you can redistribute it and/or modify
# it under the terms of the GNU Affero General Public License as published
# by the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
# cleanlab is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU Affero General Public License for more details.
# You should have received a copy of the GNU Affero General Public License
# along with cleanlab. If not, see <https://www.gnu.org/licenses/>.
Helper methods used internally for computing label quality scores
import numpy as np
from typing import Optional
from cleanlab.count import get_confident_thresholds
multi_label: bool = False,
confident_thresholds: Optional[np.ndarray] = None,
) -> np.ndarray:
"""Returns adjusted predicted probabilities by subtracting the class confident thresholds and renormalizing.
The confident class threshold for a class j is the expected (average) "self-confidence" for class j.
The purpose of this adjustment is to handle class imbalance.
labels : np.ndarray
Labels in the same format expected by the `cleanlab.count.get_confident_thresholds()` method.
If labels is None, confident_thresholds needs to be passed in as it will not be calculated.
pred_probs : np.ndarray (shape (N, K))
Predicted-probabilities in the same format expected by the `cleanlab.count.get_confident_thresholds()` method.
confident_thresholds : np.ndarray (shape (K,))
Pre-calculated confident thresholds. If passed in, function will subtract these thresholds instead of calculating
confident_thresholds from the given labels and pred_probs.
multi_label : bool, optional
If ``True``, labels should be an iterable (e.g. list) of iterables, containing a
list of labels for each example, instead of just a single label.
The multi-label setting supports classification tasks where an example has 1 or more labels.
Example of a multi-labeled `labels` input: ``[[0,1], , [0,2], [0,1,2], , , ...]``.
The major difference in how this is calibrated versus single-label is that
the total number of errors considered is based on the number of labels,
not the number of examples. So, the calibrated `confident_joint` will sum
to the number of total labels.
pred_probs_adj : np.ndarray (float)
# Get expected (average) self-confidence for each class
# TODO: Test this for multi-label
if confident_thresholds is None:
if labels is None:
f"Cannot calculate confident_thresholds without labels. Pass in either labels or already calculated "
f"confident_thresholds parameter. "
confident_thresholds = get_confident_thresholds(
labels, pred_probs, multi_label=multi_label
# Subtract the class confident thresholds
pred_probs_adj = pred_probs - confident_thresholds
# Re-normalize by shifting data to take care of negative values from the subtraction
pred_probs_adj += confident_thresholds.max()
pred_probs_adj /= pred_probs_adj.sum(axis=1)[
] # The [:, None] adds a dimension to make the /= operator work for broadcasting.
[docs]def get_normalized_entropy(pred_probs: np.ndarray, min_allowed_prob: float = 1e-6) -> np.ndarray:
"""Returns the normalized entropy of pred_probs.
Normalized entropy is between 0 and 1. Higher values of entropy indicate higher uncertainty in the model's prediction of the correct label.
Read more about normalized entropy `on Wikipedia <https://en.wikipedia.org/wiki/Entropy_(information_theory)>`_.
Normalized entropy is used in active learning for uncertainty sampling: https://towardsdatascience.com/uncertainty-sampling-cheatsheet-ec57bc067c0b
Unlike label-quality scores, entropy only depends on the model's predictions, not the given label.
Each row of this matrix corresponds to an example x and contains the model-predicted
probabilities that x belongs to each possible class: P(label=k|x)
Minimum allowed probability value. Entries of `pred_probs` below this value will be clipped to this value.
Ensures entropy remains well-behaved even when `pred_probs` contains zeros.
Each element is the normalized entropy of the corresponding row of ``pred_probs``.
num_classes = pred_probs.shape
# Note that dividing by log(num_classes) changes the base of the log which rescales entropy to 0-1 range
clipped_pred_probs = np.clip(pred_probs, a_min=min_allowed_prob, a_max=None)
return -np.sum(pred_probs * np.log(clipped_pred_probs), axis=1) / np.log(num_classes)