classification#
cleanlab can be used for learning with noisy labels for any dataset and model.
For regular (multiclass) classification tasks,
the CleanLearning
class wraps an instance of an
sklearn classifier. The wrapped classifier must adhere to the sklearn estimator API,
meaning it must define four functions:
clf.fit(X, y, sample_weight=None)
clf.predict_proba(X)
clf.predict(X)
clf.score(X, y, sample_weight=None)
where X contains data (i.e. features), y contains labels (with elements in 0, 1, …, K1,
where K is the number of classes). The first index of X and of y should correspond to the different examples in the dataset,
such that len(X) = len(y) = N
(samplesize). Here sample_weight reweights examples in
the loss function while training (supporting sample_weight in your classifier is recommended but optional).
Furthermore, your estimator should be correctly clonable via
sklearn.base.clone:
cleanlab internally creates multiple instances of the
estimator, and if you e.g. manually wrap a PyTorch model, you must ensure that
every call to the estimator’s __init__()
creates an independent instance of
the model (for sklearn compatibility, the weights of neural network models should typically be initialized inside of clf.fit()
).
Note
There are two new notions of confidence in this package:
1. Confident examples — examples we are confident are labeled correctly. We prune everything else. Mathematically, this means keeping the examples with high probability of belong to their provided label class.
2. Confident errors — examples we are confident are labeled erroneously. We prune these. Mathematically, this means pruning the examples with high probability of belong to a different class.
Examples
>>> from cleanlab.classification import CleanLearning
>>> from sklearn.linear_model import LogisticRegression as LogReg
>>> cl = CleanLearning(clf=LogReg()) # Pass in any classifier.
>>> cl.fit(X_train, labels_maybe_with_errors)
>>> # Estimate the predictions as if you had trained without label issues.
>>> pred = cl.predict(X_test)
If the model is not sklearncompatible by default, it might be the case that standard packages can adapt the model. For example, you can adapt PyTorch models using skorch and adapt Keras models using SciKeras.
If an opensource adapter doesn’t already exist, you can manually wrap the model to be sklearncompatible. This is made easy by inheriting from sklearn.base.BaseEstimator:
from sklearn.base import BaseEstimator
class YourModel(BaseEstimator):
def __init__(self, ):
pass
def fit(self, X, y, sample_weight=None):
pass
def predict(self, X):
pass
def predict_proba(self, X):
pass
def score(self, X, y, sample_weight=None):
pass
Note
labels refers to the given labels in the original dataset, which may have errors
labels must be integers in 0, 1, …, K1, where K is the total number of classes
Note
Confident learning is the stateoftheart (Northcutt et al., 2021) for weak supervision, finding label issues in datasets, learning with noisy labels, uncertainty estimation, and more. It works with any classifier, including deep neural networks. See the clf parameter.
Confident learning is a subfield of theory and algorithms of machine learning with noisy labels. Cleanlab achieves stateoftheart performance of any opensourced implementation of confident learning across a variety of tasks like multiclass classification, multilabel classification, and PU learning.
Given any classifier having the predict_proba method, an input feature matrix X, and a discrete vector of noisy labels labels, confident learning estimates the classifications that would be obtained if the true labels had instead been provided to the classifier during training. labels denotes the noisy labels instead of the $\tilde{y}$ used in confident learning paper.
Classes:

CleanLearning = Machine Learning with cleaned data (even when training on messy, errorridden data). 
 class cleanlab.classification.CleanLearning(clf=None, *, seed=None, cv_n_folds=5, converge_latent_estimates=False, pulearning=None, find_label_issues_kwargs={}, label_quality_scores_kwargs={}, verbose=False)[source]#
Bases:
BaseEstimator
CleanLearning = Machine Learning with cleaned data (even when training on messy, errorridden data).
Automated and robust learning with noisy labels using any dataset and any model. This class trains a model clf with errorprone, noisy labels as if the model had been instead trained on a dataset with perfect labels. It achieves this by cleaning out the error and providing cleaned data while training. This class is currently intended for standard (multiclass) classification tasks.
 Parameters:
clf (
estimator instance
, optional) –A classifier implementing the sklearn estimator API, defining the following functions:
clf.fit(X, y, sample_weight=None)
clf.predict_proba(X)
clf.predict(X)
clf.score(X, y, sample_weight=None)
See
cleanlab.experimental
for examples of sklearn wrappers, e.g. around PyTorch and FastText.If the model is not sklearncompatible by default, it might be the case that standard packages can adapt the model. For example, you can adapt PyTorch models using skorch and adapt Keras models using SciKeras.
Stores the classifier used in Confident Learning. Default classifier used is sklearn.linear_model.LogisticRegression.
seed (
int
, optional) – Set the default state of the random number generator used to split the crossvalidated folds. By default, uses np.random current random state.cv_n_folds (
int
, default5
) – This class needs holdout predicted probabilities for every data example and if not provided, uses crossvalidation to compute them. cv_n_folds sets the number of crossvalidation folds used to compute outofsample probabilities for each example in X.converge_latent_estimates (
bool
, optional) – If true, forces numerical consistency of latent estimates. Each is estimated independently, but they are related mathematically with closed form equivalences. This will iteratively enforce consistency.pulearning (
{None, 0, 1}
, defaultNone
) – Only works for 2 class datasets. Set to the integer of the class that is perfectly labeled (you are certain that there are no errors in that class).find_label_issues_kwargs (
dict
, optional) – Keyword arguments to pass intofilter.find_label_issues
. Options that may especially impact accuracy include: filter_by, frac_noise, min_examples_per_class.label_quality_scores_kwargs (
dict
, optional) – Keyword arguments to pass intorank.get_label_quality_scores
. Options include: method, adjust_pred_probs.verbose (
bool
, defaultFalse
) – Controls how much output is printed. Set toFalse
to suppress print statements.
Methods:
find_label_issues
([X, labels, pred_probs, ...])Identifies potential label issues in the dataset using confident learning.
fit
(X[, labels, pred_probs, thresholds, ...])Train the model clf with errorprone, noisy labels as if the model had been instead trained on a dataset with the correct labels.
Accessor.
get_params
([deep])Get parameters for this estimator.
predict
(*args, **kwargs)Predict class labels using your wrapped classifier clf.
predict_proba
(*args, **kwargs)Predict class probabilities
P(true label=k)
using your wrapped classifier clf.Clears nonsklearn attributes of this estimator to save space (inplace).
score
(X, y[, sample_weight])Evaluates your wrapped classifier clf's score on a test set X with labels y.
set_params
(**params)Set the parameters of this estimator.
 find_label_issues(X=None, labels=None, *, pred_probs=None, thresholds=None, noise_matrix=None, inverse_noise_matrix=None, save_space=False, clf_kwargs={}, validation_func=None)[source]#
Identifies potential label issues in the dataset using confident learning.
Runs crossvalidation to get outofsample pred_probs from clf and then calls
filter.find_label_issues
to find label issues. These label issues are cached internally and returned in a pandas DataFrame. Kwargs forfilter.find_label_issues
must have already been specified in the initialization of this class, not here.Unlike
filter.find_label_issues
, which requires pred_probs, this method only requires a classifier and it can do the crossvalidation for you. Both methods return the same boolean mask that identifies which examples have label issues. This is the preferred method to use if you plan to subsequently invoke:CleanLearning.fit()
.Note: this method computes the label issues from scratch. To access previouslycomputed label issues from this
CleanLearning
instance, use theget_label_issues
method.This is the method called to find label issues inside
CleanLearning.fit()
and they share mostly the same parameters. Parameters:
save_space (
bool
, optional) – If True, then returned label_issues_df will not be stored as attribute. This means some other methods like self.get_label_issues() will no longer work.
For info about the other parameters, see the docstring of
CleanLearning.fit()
. Return type:
DataFrame
 Returns:
label_issues_df (
pd.DataFrame
) – DataFrame with info about label issues for each example. Unless save_space argument is specified, same DataFrame is also stored as self.label_issues_df attribute accessible viaget_label_issues
. Each row represents an example from our dataset and the DataFrame may contain the following columns:is_label_issue: boolean mask for the entire dataset where
True
represents a label issue andFalse
represents an example that is accurately labeled with high confidence. This column is equivalent to label_issues_mask output fromfilter.find_label_issues
.label_quality: Numeric score that measures the quality of each label (how likely it is to be correct, with lower scores indicating potentially erroneous labels).
given_label: Integer indices corresponding to the class label originally given for this example (same as labels input). Included here for ease of comparison against clf predictions, only present if “predicted_label” column is present.
predicted_label: Integer indices corresponding to the class predicted by trained clf model. Only present if
pred_probs
were provided as input or computed during labelissuefinding.sample_weight: Numeric values used to weight examples during the final training of clf in
CleanLearning.fit()
. This column may not be present after self.find_label_issues() but may be added after call toCleanLearning.fit()
. For more precise definition of sample weights, see documentation ofCleanLearning.fit()
 fit(X, labels=None, *, pred_probs=None, thresholds=None, noise_matrix=None, inverse_noise_matrix=None, label_issues=None, sample_weight=None, clf_kwargs={}, clf_final_kwargs={}, validation_func=None, y=None)[source]#
Train the model clf with errorprone, noisy labels as if the model had been instead trained on a dataset with the correct labels. fit achieves this by first training clf via crossvalidation on the noisy data, using the resulting predicted probabilities to identify label issues, pruning the data with label issues, and finally training clf on the remaining clean data.
 Parameters:
X (
np.ndarray
orDatasetLike
) –Data features (i.e. training inputs for ML), typically an array of shape
(N, ...)
, where N is the number of examples. Supported DatasetLike types beyondnp.ndarray
include:pd.DataFrame
,scipy.sparse.csr_matrix
,torch.utils.data.Dataset
,tensorflow.data.Dataset
, or any dataset objectX
that supports listbased indexing:X[index_list]
to select a subset of training examples. Your classifier that this instance was initialized with,clf
, must be able tofit()
andpredict()
data of this format.Note
If providing X as a
tensorflow.data.Dataset
, make sureshuffle()
has been called beforebatch()
(if shuffling) and no other orderdestroying operation (eg.repeat()
) has been applied.labels (
array_like
) – An array of shape(N,)
of noisy classification labels, where some labels may be erroneous. Elements must be integers in the set 0, 1, …, K1, where K is the number of classes. Supported array_like types include:np.ndarray
,pd.Series
, orlist
.pred_probs (
np.ndarray
, optional) –An array of shape
(N, K)
of modelpredicted probabilities,P(label=kx)
. Each row of this matrix corresponds to an example x and contains the modelpredicted probabilities that x belongs to each possible class, for each of the K classes. The columns must be ordered such that these probabilities correspond to class 0, 1, …, K1. pred_probs should be outofsample, eg. computed via crossvalidation. If provided, pred_probs will be used to find label issues rather than theclf
classifier.Note
If you are not sure, leave
pred_probs=None
(the default) and it will be computed for you using crossvalidation with the provided model.thresholds (
array_like
, optional) –An array of shape
(K, 1)
or(K,)
of perclass threshold probabilities, used to determine the cutoff probability necessary to consider an example as a given class label (see Northcutt et al., 2021, Section 3.1, Equation 2).This is for advanced users only. If not specified, these are computed for you automatically. If an example has a predicted probability greater than this threshold, it is counted as having true_label = k. This is not used for pruning/filtering, only for estimating the noise rates using confident counts.
noise_matrix (
np.ndarray
, optional) – An array of shape(K, K)
representing the conditional probability matrixP(label=k_s  true label=k_y)
, the fraction of examples in every class, labeled as every other class. Assumes columns of noise_matrix sum to 1.inverse_noise_matrix (
np.ndarray
, optional) – An array of shape(K, K)
representing the conditional probability matrixP(true label=k_y  label=k_s)
, the estimated fraction observed examples in each classk_s
that are mislabeled examples from every other classk_y
, Assumes columns of inverse_noise_matrix sum to 1.label_issues (
pd.DataFrame
ornp.ndarray
, optional) –Specifies the label issues for each example in dataset. If
pd.DataFrame
, must be formatted as the one returned by:CleanLearning.find_label_issues
orCleanLearning.get_label_issues
. Ifnp.ndarray
, must contain either boolean label_issues_mask as output by: defaultfilter.find_label_issues
, or integer indices as output byfilter.find_label_issues
with its return_indices_ranked_by argument specified. Providing this argument significantly reduces the time this method takes to run by skipping the slow crossvalidation step necessary to find label issues. Examples identified to have label issues will be pruned from the data before training the final clf model.Caution: If you provide label_issues without having previously called
self.find_label_issues
, e.g. as anp.ndarray
, then some functionality like training with sample weights may be disabled.sample_weight (
array_like
, optional) – Array of weights with shape(N,)
that are assigned to individual samples, assuming total number of examples in dataset is N. If not provided, samples may still be weighted by the estimated noise in the class they are labeled as.clf_kwargs (
dict
, optional) – Optional keyword arguments to pass into clf’sfit()
method.clf_final_kwargs (
dict
, optional) – Optional extra keyword arguments to pass into the final clffit()
on the cleaned data but not the clffit()
in each fold of crossvalidation on the noisy data. The finalfit()
will also receive clf_kwargs, but these may be overwritten by values in clf_final_kwargs. This can be useful for training differently in the finalfit()
than during crossvalidation.validation_func (
callable
, optional) –Optional callable function that takes two arguments, X_val, y_val, and returns a dict of keyword arguments passed into to
clf.fit()
which may be functions of the validation data in each crossvalidation fold. Specifies how to map the validation data split in each crossvalidation fold into the appropriate format to pass into clf’sfit()
method, assumingclf.fit()
can utilize validation data if it is appropriately passed in (eg. for earlystopping). Eg. if your model’sfit()
method is called usingclf.fit(X, y, X_validation, y_validation)
, then you could setvalidation_func = f
wheredef f(X_val, y_val): return {"X_validation": X_val, "y_validation": y_val}
Note that validation_func will be ignored in the final call to clf.fit() on the cleaned subset of the data. This argument is only for allowing clf to access the validation data in each crossvalidation fold (eg. for earlystopping or hyperparameterselection purposes). If you want to pass in validation data even in the final training call to
clf.fit()
on the cleaned data subset, you should explicitly pass in that data yourself (eg. via clf_final_kwargs or clf_kwargs).y (
array_like
, optional) – Alternative argument that can be specified instead of labels. Specifying y has the same effect as specifying labels, and is offered as an alternative for compatibility with sklearn.
 Return type:
TypeVar
(TCleanLearning
, bound= CleanLearning) Returns:
self (
CleanLearning
) – Fitted estimator that has all the same methods as any sklearn estimator.After calling
self.fit()
, this estimator also stores extra attributes such as:self.label_issues_df: a
pd.DataFrame
accessible via
get_label_issues
of similar format as the one returned by:CleanLearning.find_label_issues
. See documentation ofCleanLearning.find_label_issues
for column descriptions.After calling
self.fit()
, self.label_issues_df may also contain an extra column:sample_weight: Numeric values that were used to weight examples during the final training of clf in
CleanLearning.fit()
. sample_weight column will only be present if automatic sample weights were actually used. These automatic weights are assigned to each example based on the class it belongs to, i.e. there are only num_classes unique sample_weight values. The sample weight for an example belonging to class k is computed as1 / p(given_label = k  true_label = k)
. This sample_weight normalizes the loss to effectively trick clf into learning with the distribution of the true labels by accounting for the noisy data pruned out prior to training on cleaned data. In other words, examples with label issues were removed, so this weights the data proportionally so that the classifier trains as if it had all the true labels, not just the subset of cleaned data left after pruning out the label issues.
Note
If
CleanLearning.fit()
does not work for your data/model, you can run the same procedure yourself: * Utilize crossvalidation to get outofsample pred_probs for each example. * Callfilter.find_label_issues
with pred_probs. * Filter the examples with detected issues and train your model on the remaining data.
 get_label_issues()[source]#
Accessor. Returns label_issues_df attribute if previously already computed. This
pd.DataFrame
describes the label issues identified for each example (each row corresponds to an example). For column definitions, see the documentation ofCleanLearning.find_label_issues
. Return type:
Optional
[DataFrame
] Returns:
label_issues_df (
pd.DataFrame
) – DataFrame with (precomputed) info about label issues for each example.
 get_params(deep=True)#
Get parameters for this estimator.
 Parameters:
deep (
bool
, defaultTrue
) – If True, will return the parameters for this estimator and contained subobjects that are estimators. Returns:
params (
dict
) – Parameter names mapped to their values.
 predict(*args, **kwargs)[source]#
Predict class labels using your wrapped classifier clf. Works just like
clf.predict()
. Parameters:
X (
np.ndarray
orDatasetLike
) – Test data in the same format expected by your wrapped classifier. Return type:
ndarray
 Returns:
class_predictions (
np.ndarray
) – Vector of class predictions for the test examples.
 predict_proba(*args, **kwargs)[source]#
Predict class probabilities
P(true label=k)
using your wrapped classifier clf. Works just likeclf.predict_proba()
. Parameters:
X (
np.ndarray
orDatasetLike
) – Test data in the same format expected by your wrapped classifier. Return type:
ndarray
 Returns:
pred_probs (
np.ndarray
) –(N x K)
array of predicted class probabilities, one row for each test example.
 save_space()[source]#
Clears nonsklearn attributes of this estimator to save space (inplace). This includes the DataFrame attribute that stored label issues which may be large for big datasets. You may want to call this method before deploying this model (i.e. if you just care about producing predictions). After calling this method, certain nonpredictionrelated attributes/functionality will no longer be available (e.g. you cannot call
self.fit()
anymore).
 score(X, y, sample_weight=None)[source]#
Evaluates your wrapped classifier clf’s score on a test set X with labels y. Uses your model’s default scoring function, or simply accuracy if your model as no
"score"
attribute. Parameters:
X (
np.ndarray
orDatasetLike
) – Test data in the same format expected by your wrapped classifier.y (
array_like
) – Test labels in the same format as labels previously used infit()
.sample_weight (
np.ndarray
, optional) – An array of shape(N,)
or(N, 1)
used to weight each test example when computing the score.
 Return type:
float
 Returns:
score (
float
) – Number quantifying the performance of this classifier on the test data.
 set_params(**params)#
Set the parameters of this estimator.
The method works on simple estimators as well as on nested objects (such as
Pipeline
). The latter have parameters of the form<component>__<parameter>
so that it’s possible to update each component of a nested object. Parameters:
**params (
dict
) – Estimator parameters. Returns:
self (
estimator instance
) – Estimator instance.