cleanlab is a machine learning python package for learning with noisy labels and finding label errors in datasets. cleanlab CLEANs LABels.

cleanlab finds and cleans label errors in any dataset using state-of-the-art algorithms for learning with noisy labels and estimating the complete joint distribution of label noise. cleanlab is fast: its built on optimized algorithms and parallelized across all CPU threads automatically. cleanlab implements the family of theory and algorithms called confident learning with provable guarantees of exact noise estimation and label error finding (when model output probabilities are exact).

Its called cleanlab because it CLEANs LABels.

cleanlab is:

1. fast - runs in parallel on all CPU threads automatically. (e.g. < 1 second to find label errors in ImageNet)
2. robust - Provable generalization and risk minimimzation guarantees, including imperfect probability estimation.
3. general - Works with any probablistic classifier: PyTorch, Tensorflow, MxNet, Caffe2, scikit-learn, etc.
4. unique - The only package for multiclass learning with noisy labels or finding label errors for any dataset / classifier.

from cleanlab.classification import LearningWithNoisyLabels
from sklearn.linear_model import LogisticRegression

# Learning with noisy labels in 3 lines of code!

# Wrap around any classifier. Yup, you can use sklearn/pyTorch/Tensorflow/FastText/etc.
lnl = LearningWithNoisyLabels(clf=LogisticRegression())
lnl.fit(X = X_train_data, s = train_noisy_labels)
# Estimate the predictions you would have gotten by training with *no* label errors.
predicted_test_labels = lnl.predict(X_test)

Check out these examples and tests (includes how to use pyTorch, FastText, etc.).

Installation

Python 2.7, 3.4, 3.5, and 3.6 are supported.

Stable release:

$ pip install cleanlab

Developer (unstable) release:

$ pip install git+https://github.com/cgnorthcutt/cleanlab.git

To install the codebase (enabling you to make modifications):

$ conda update pip # if you use conda
$ git clone https://github.com/cgnorthcutt/cleanlab.git
$ cd cleanlab
$ pip install -e .

Citations and Related Publications

Although this package goes far beyond our 2017 publication, if you find this repository helpful, please cite our paper http://auai.org/uai2017/proceedings/papers/35.pdf. New papers will be posted here when they are published.

@inproceedings{northcutt2017rankpruning,
 author={Northcutt, Curtis G. and Wu, Tailin and Chuang, Isaac L.},
 title={Learning with Confident Examples: Rank Pruning for Robust Classification with Noisy Labels},
 booktitle = {Proceedings of the Thirty-Third Conference on Uncertainty in Artificial Intelligence},
 series = {UAI'17},
 year = {2017},
 location = {Sydney, Australia},
 numpages = {10},
 url = {http://auai.org/uai2017/proceedings/papers/35.pdf},
 publisher = {AUAI Press},
}

Collaboration

Most of the algorithms, theory, and results of cleanlab remain unpublished. If you'd like to work together, please reach out.

cleanlab on MNIST

We use cleanlab to automatically identify ~50 label errors in the MNIST dataset.

Label errors of the original MNIST train dataset identified algorithmically using the rankpruning algorithm. Depicts the 24 least confident labels, ordered left-right, top-down by increasing self-confidence (probability of belonging to the given label), denoted conf in teal. The label with the largest predicted probability is in green. Overt errors are in red.

cleanlab Generality: View performance across 4 distributions and 9 classifiers.

We use cleanlab to automatically learn with noisy labels regardless of dataset distribution or classifier.

Each figure depicts the decision boundary learned using cleanlab.classification.LearningWithNoisyLabels in the presence of extreme (~35%) label errors. Label errors are circled in green. Label noise is class-conditional (not simply uniformly random). Columns are organized by the classifier used, except the left-most column which depicts the ground-truth dataset distribution. Rows are organized by dataset used. A matrix characterizing the label noise for the first row is shown below.

Each figure depicts accuracy scores on a test set as decimal values:

1. LEFT (in black): The classifier test accuracy trained with perfect labels (no label errors).
2. MIDDLE (in blue): The classifier test accuracy trained with noisy labels using cleanlab.
3. RIGHT (in white): The baseline classifier test accuracy trained with noisy labels.

As an example, this is the noise matrix (noisy channel) P(s | y) characterizing the label noise for the first dataset row in the figure. s represents the observed noisy labels and y represents the latent, true labels. The trace of this matrix is 2.6. A trace of 4 implies no label noise. A cell in this matrix is read like, "A random 38% of '3' labels were flipped to '2' labels."

p(s|y)	y=0	y=1	y=2	y=3
s=0	0.55	0.01	0.07	0.06
s=1	0.22	0.87	0.24	0.02
s=2	0.12	0.04	0.64	0.38
s=3	0.11	0.08	0.05	0.54

The code to reproduce this figure is available here.

Get started with easy, quick examples.

New to cleanlab? Start with:

1. Visualizing confident learning
2. A simple example of learning with noisy labels on the multiclass Iris dataset.

These examples show how easy it is to characterize label noise in datasets, learn with noisy labels, identify label errors, estimate latent priors and noisy channels, and more.

Use cleanlab with any model (Tensorflow, caffe2, PyTorch, etc.)

All of the features of the cleanlab package work with any model. Yes, any model. Feel free to use PyTorch, Tensorflow, caffe2, scikit-learn, mxnet, etc. If you use a scikit-learn classifier, all cleanlab methods will work out-of-the-box. It’s also easy to use your favorite model from a non-scikit-learn package, just wrap your model into a Python class that inherits the sklearn.base.BaseEstimator:

from sklearn.base import BaseEstimator
class YourFavoriteModel(BaseEstimator): # Inherits sklearn base classifier
    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

# Now you can use your model with `cleanlab`. Here's one example:
from cleanlab.classification import LearningWithNoisyLabels
lnl = LearningWithNoisyLabels(clf=YourFavoriteModel())
lnl.fit(train_data, train_labels_with_errors)

Want to see a working example? Here’s a compliant PyTorch MNIST CNN class

As you can see here, technically you don’t actually need to inherit from sklearn.base.BaseEstimator, as you can just create a class that defines .fit(), .predict(), and .predict_proba(), but inheriting makes downstream scikit-learn applications like hyper-parameter optimization work seamlessly. For example, the LearningWithNoisyLabels() model is fully compliant.

Note, some libraries exists to do this for you. For pyTorch, check out the skorch Python library which will wrap your pytorch model into a scikit-learn compliant model.

Documentation by Example

cleanlab Core Package Components

1. cleanlab/classification.py - The LearningWithNoisyLabels() class for learning with noisy labels.
2. cleanlab/latent_algebra.py - Equalities when noise information is known.
3. cleanlab/latent_estimation.py - Estimates and fully characterizes all variants of label noise.
4. cleanlab/noise_generation.py - Generate mathematically valid synthetic noise matrices.
5. cleanlab/polyplex.py - Characterizes joint distribution of label noise EXACTLY from noisy channel.
6. cleanlab/pruning.py - Finds the indices of the examples with label errors in a dataset.

Many of these methods have default parameters that won’t be covered here. Check out the method docstrings for full documentation.

Multiclass learning with noisy labels (in 3 lines of code):

rankpruning is a fast, general, robust algorithm for multiclass learning with noisy labels. It adds minimal overhead, needing only O(nm2) time for n training examples and m classes, works with any classifier, and is easy to use. Here is the example from above, with added commments for clarity.

# LearningWithNoisyLabels implements a faster,
# cross-platform and more-compatible version of the RankPruning
# algorithm for learning with noisy labels. Unlike the original
# algorithm which only worked for binary classification,
# LearningWithNoisyLabels generalizes the theory and algorithms
# of RankPruning for any number of classes.
from cleanlab.classification import LearningWithNoisyLabels
# LearningWithNoisyLabels uses logreg by default, so this is unnecessary.
# We include it here for clarity, but this step is omitted below.
from sklearn.linear_model import LogisticRegression as logreg

# 1.
# Wrap around any classifier. Yup, neural networks work, too.
lnl = LearningWithNoisyLabels(clf=logreg())

# 2.
# X_train is numpy matrix of training examples (integers for large data)
# train_labels_with_errors is a numpy array of labels of length n (# of examples), usually denoted 's'.
lnl.fit(X_train, train_labels_with_errors)

# 3.
# Estimate the predictions you would have gotten by training with *no* label errors.
predicted_test_labels = lnl.predict(X_test)

Estimate the confident joint, the latent noisy channel matrix, P(s | y) and inverse, P(y | s), the latent prior of the unobserved, actual true labels, p(y), and the predicted probabilities.:

where s denotes a random variable that represents the observed, noisy label and y denotes a random variable representing the hidden, actual labels. Both s and y take any of the m classes as values. The cleanlab package supports different levels of granularity for computation depending on the needs of the user. Because of this, we support multiple alternatives, all no more than a few lines, to estimate these latent distribution arrays, enabling the user to reduce computation time by only computing what they need to compute, as seen in the examples below.

Throughout these examples, you’ll see a variable called confident_joint. The confident joint is an m x m matrix (m is the number of classes) that counts, for every observed, noisy class, the number of examples that confidently belong to every latent, hidden class. It counts the number of examples that we are confident are labeled correctly or incorrectly for every pair of obseved and unobserved classes. The confident joint is an unnormalized estimate of the complete-information latent joint distribution, Ps,y. Most of the methods in the cleanlab package start by first estimating the confident_joint.

Option 1: Compute the confident joint and predicted probs first. Stop if that’s all you need.

from cleanlab.latent_estimation import estimate_latent
from cleanlab.latent_estimation import estimate_confident_joint_and_cv_pred_proba

# Compute the confident joint and the n x m predicted probabilities matrix (psx),
# for n examples, m classes. Stop here if all you need is the confident joint.
confident_joint, psx = estimate_confident_joint_and_cv_pred_proba(
    X=X_train,
    s=train_labels_with_errors,
    clf = logreg(), # default, you can use any classifier
)

# Estimate latent distributions: p(y) as est_py, P(s|y) as est_nm, and P(y|s) as est_inv
est_py, est_nm, est_inv = estimate_latent(confident_joint, s=train_labels_with_errors)

Option 2: Estimate the latent distribution matrices in a single line of code.

from cleanlab.latent_estimation import estimate_py_noise_matrices_and_cv_pred_proba
est_py, est_nm, est_inv, confident_joint, psx = estimate_py_noise_matrices_and_cv_pred_proba(
    X=X_train,
    s=train_labels_with_errors,
)

Option 3: Skip computing the predicted probabilities if you already have them.

# Already have psx? (n x m matrix of predicted probabilities)
# For example, you might get them from a pre-trained model (like resnet on ImageNet)
# With the cleanlab package, you estimate directly with psx.
from cleanlab.latent_estimation import estimate_py_and_noise_matrices_from_probabilities
est_py, est_nm, est_inv, confident_joint = estimate_py_and_noise_matrices_from_probabilities(
    s=train_labels_with_errors,
    psx=psx,
)

Estimate label errors in a dataset:

With the cleanlab package, we can instantly fetch the indices of all estimated label errors, with nothing provided by the user except a classifier, examples, and their noisy labels. Like the previous example, there are various levels of granularity.

from cleanlab.pruning import get_noise_indices
# We computed psx, est_inv, confident_joint in the previous example.
label_errors = get_noise_indices(
    s=train_labels_with_errors, # required
    psx=psx, # required
    inverse_noise_matrix=est_inv, # not required, include to avoid recomputing
    confident_joint=confident_joint, # not required, include to avoid recomputing
)

Estimate the latent joint probability distribution matrix of the noisy and true labels, P(s,y):

To compute P(s,y), the complete-information distribution matrix that captures the number of pairwise label flip errors when multipled by the total number of examples as n P(s,y)*. Using cleanlab.latent_estimation.calibrate_confident_joint, this method guarantees the rows of P(s,y) correctly sum to p(s), and the entire matrix sums to 1.

This method occurs when hyperparameter prune_count_method = ‘inverse_nm_dot_s’ in LearningWithNoisyLabels.fit() and get_noise_indices().

from cleanlab.latent_estimation import estimate_confident_joint_from_probabilities
joint = estimate_confident_joint_from_probabilities(s=noisy_labels, psx=probabilities)

If you already have the confident joint, then you can quickly estimate the complete joint distribution of label noise by:

from cleanlab.latent_estimation import estimate_joint
joint = estimate_joint(confident_joint=cj, s=noisy_labels, psx=probabilities)

Generate valid, class-conditional, unformly random noisy channel matrices:

# Generate a valid (necessary conditions for learnability are met) noise matrix for any trace > 1
from cleanlab.noise_generation import generate_noise_matrix_from_trace
noise_matrix = generate_noise_matrix_from_trace(
    K = number_of_classes,
    trace = float_value_greater_than_1_and_leq_K,
    py = prior_of_y_actual_labels_which_is_just_an_array_of_length_K,
    frac_zero_noise_rates = float_from_0_to_1_controlling_sparsity,
)

# Check if a noise matrix is valid (necessary conditions for learnability are met)
from cleanlab.noise_generation import noise_matrix_is_valid
is_valid = noise_matrix_is_valid(noise_matrix, prior_of_y_which_is_just_an_array_of_length_K)

Support for numerous weak supervision and learning with noisy labels functionalities:

# Generate noisy labels using the noise_marix. Guarantees exact amount of noise in labels.
from cleanlab.noise_generation import generate_noisy_labels
s_noisy_labels = generate_noisy_labels(y_hidden_actual_labels, noise_matrix)

# This package is a full of other useful methods for learning with noisy labels.
# The tutorial stops here, but you don't have to. Inspect method docstrings for full docs.

The Polyplex

The key to learning in the presence of label errors is estimating the joint distribution between the actual, hidden labels ‘y’ and the observed, noisy labels ‘s’. Using cleanlab and the theory of confident learning, we can completely characterize the trace of the latent joint distribution, trace(P(s,y)), given p(y), for any fraction of label errors, i.e. for any trace of the noisy channel, trace(P(s|y)).

You can check out how to do this yourself here: 1. Drawing Polyplices 2. Computing Polyplices

License

Copyright (c) 2017-2019 Curtis Northcutt. Released under the MIT License. See LICENSE for details.