Source code for cleanlab.internal.util

# Copyright (C) 2017-2023  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/>.

"""
Ancillary helper methods used internally throughout this package; mostly related to Confident Learning algorithms.
"""

import warnings
import numpy as np
import pandas as pd
from typing import Union, Tuple

from cleanlab.typing import DatasetLike, LabelLike
from cleanlab.internal.validation import labels_to_array
from cleanlab.internal.constants import FLOATING_POINT_COMPARISON, TINY_VALUE


[docs]def remove_noise_from_class(noise_matrix, class_without_noise) -> np.ndarray: """A helper function in the setting of PU learning. Sets all P(label=class_without_noise|true_label=any_other_class) = 0 in noise_matrix for pulearning setting, where we have generalized the positive class in PU learning to be any class of choosing, denoted by class_without_noise. Parameters ---------- noise_matrix : np.ndarray of shape (K, K), K = number of classes A conditional probability matrix of the form P(label=k_s|true_label=k_y) containing the fraction of examples in every class, labeled as every other class. Assumes columns of noise_matrix sum to 1. class_without_noise : int Integer value of the class that has no noise. Traditionally, this is 1 (positive) for PU learning.""" # Number of classes K = len(noise_matrix) cwn = class_without_noise x = np.copy(noise_matrix) # Set P( labels = cwn | y != cwn) = 0 (no noise) x[cwn, [i for i in range(K) if i != cwn]] = 0.0 # Normalize columns by increasing diagonal terms # Ensures noise_matrix is a valid probability matrix for i in range(K): x[i][i] = 1 - float(np.sum(x[:, i]) - x[i][i]) return x
[docs]def clip_noise_rates(noise_matrix) -> np.ndarray: """Clip all noise rates to proper range [0,1), but do not modify the diagonal terms because they are not noise rates. ASSUMES noise_matrix columns sum to 1. Parameters ---------- noise_matrix : np.ndarray of shape (K, K), K = number of classes A conditional probability matrix containing the fraction of examples in every class, labeled as every other class. Diagonal terms are not noise rates, but are consistency P(label=k|true_label=k) Assumes columns of noise_matrix sum to 1""" def clip_noise_rate_range(noise_rate) -> float: """Clip noise rate P(label=k'|true_label=k) or P(true_label=k|label=k') into proper range [0,1)""" return min(max(noise_rate, 0.0), 0.9999) # Vectorize clip_noise_rate_range for efficiency with np.ndarrays. vectorized_clip = np.vectorize(clip_noise_rate_range) # Preserve because diagonal entries are not noise rates. diagonal = np.diagonal(noise_matrix) # Clip all noise rates (efficiently). noise_matrix = vectorized_clip(noise_matrix) # Put unmodified diagonal back. np.fill_diagonal(noise_matrix, diagonal) # Re-normalized noise_matrix so that columns sum to one. noise_matrix = noise_matrix / np.clip(noise_matrix.sum(axis=0), a_min=TINY_VALUE, a_max=None) return noise_matrix
[docs]def clip_values(x, low=0.0, high=1.0, new_sum=None) -> np.ndarray: """Clip all values in p to range [low,high]. Preserves sum of x. Parameters ---------- x : np.ndarray An array / list of values to be clipped. low : float values in x greater than 'low' are clipped to this value high : float values in x greater than 'high' are clipped to this value new_sum : float normalizes x after clipping to sum to new_sum Returns ------- x : np.ndarray A list of clipped values, summing to the same sum as x.""" def clip_range(a, low=low, high=high): """Clip a into range [low,high]""" return min(max(a, low), high) vectorized_clip = np.vectorize( clip_range ) # Vectorize clip_range for efficiency with np.ndarrays prev_sum = sum(x) if new_sum is None else new_sum # Store previous sum x = vectorized_clip(x) # Clip all values (efficiently) x = ( x * prev_sum / np.clip(float(sum(x)), a_min=TINY_VALUE, a_max=None) ) # Re-normalized values to sum to previous sum return x
[docs]def value_counts(x, *, num_classes=None, multi_label=False) -> np.ndarray: """Returns an np.ndarray of shape (K, 1), with the value counts for every unique item in the labels list/array, where K is the number of unique entries in labels. Works for both single-labeled and multi-labeled data. Parameters ---------- x : list or np.ndarray (one dimensional) A list of discrete objects, like lists or strings, for example, class labels 'y' when training a classifier. e.g. ["dog","dog","cat"] or [1,2,0,1,1,0,2] num_classes : int (default: None) Setting this fills the value counts for missing classes with zeros. For example, if x = [0, 0, 1, 1, 3] then setting ``num_classes=5`` returns [2, 2, 0, 1, 0] whereas setting ``num_classes=None`` would return [2, 2, 1]. This assumes your labels come from the set [0, 1,... num_classes=1] even if some classes are missing. 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. Assumes all classes in pred_probs.shape[1] are represented in labels. The multi-label setting supports classification tasks where an example has 1 or more labels. Example of a multi-labeled `labels` input: ``[[0,1], [1], [0,2], [0,1,2], [0], [1], ...]``. 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.""" # Efficient method if x is pd.Series, np.ndarray, or list if multi_label: x = [z for lst in x for z in lst] # Flatten unique_classes, counts = np.unique(x, return_counts=True) if num_classes is None or num_classes == len(unique_classes): return counts # Else, there are missing classes if num_classes <= max(unique_classes): raise ValueError(f"Required: num_classes > max(x), but {num_classes} <= {max(x)}.") # Add zero counts for all missing classes in [0, 1,..., num_classes-1] # multi_label=False regardless because x was flattened. missing_classes = get_missing_classes(x, num_classes=num_classes, multi_label=False) missing_counts = [(z, 0) for z in missing_classes] # Return counts with zeros for all missing classes. return np.array(list(zip(*sorted(list(zip(unique_classes, counts)) + missing_counts)))[1])
[docs]def value_counts_fill_missing_classes(x, num_classes, *, multi_label=False) -> np.ndarray: """Same as ``internal.util.value_counts`` but requires that num_classes is provided and always fills missing classes with zero counts. See ``internal.util.value_counts`` for parameter docstrings.""" return value_counts(x, num_classes=num_classes, multi_label=multi_label)
[docs]def get_missing_classes(labels, *, pred_probs=None, num_classes=None, multi_label=False): """Find which classes are present in ``pred_probs`` but not present in ``labels``. See ``count.compute_confident_joint`` for parameter docstrings.""" if pred_probs is None and num_classes is None: raise ValueError("Both pred_probs and num_classes are None. You must provide exactly one.") if pred_probs is not None and num_classes is not None: raise ValueError("Both pred_probs and num_classes are not None. Only one may be provided.") if num_classes is None: num_classes = pred_probs.shape[1] unique_classes = get_unique_classes(labels, multi_label=multi_label) return sorted(set(range(num_classes)).difference(unique_classes))
[docs]def round_preserving_sum(iterable) -> np.ndarray: """Rounds an iterable of floats while retaining the original summed value. The name of each parameter is required. The type and description of each parameter is optional, but should be included if not obvious. The while loop in this code was adapted from: https://github.com/cgdeboer/iteround Parameters ----------- iterable : list<float> or np.ndarray<float> An iterable of floats Returns ------- list<int> or np.ndarray<int> The iterable rounded to int, preserving sum.""" floats = np.asarray(iterable, dtype=float) ints = floats.round() orig_sum = np.sum(floats).round() int_sum = np.sum(ints).round() # Adjust the integers so that they sum to orig_sum while abs(int_sum - orig_sum) > FLOATING_POINT_COMPARISON: diff = np.round(orig_sum - int_sum) increment = -1 if int(diff < 0.0) else 1 changes = min(int(abs(diff)), len(iterable)) # Orders indices by difference. Increments # of changes. indices = np.argsort(floats - ints)[::-increment][:changes] for i in indices: ints[i] = ints[i] + increment int_sum = np.sum(ints).round() return ints.astype(int)
[docs]def round_preserving_row_totals(confident_joint) -> np.ndarray: """Rounds confident_joint cj to type int while preserving the totals of reach row. Assumes that cj is a 2D np.ndarray of type float. Parameters ---------- confident_joint : 2D np.ndarray<float> of shape (K, K) See compute_confident_joint docstring for details. Returns ------- confident_joint : 2D np.ndarray<int> of shape (K,K) Rounded to int while preserving row totals.""" return np.apply_along_axis( func1d=round_preserving_sum, axis=1, arr=confident_joint, ).astype(int)
[docs]def estimate_pu_f1(s, prob_s_eq_1) -> float: """Computes Claesen's estimate of f1 in the pulearning setting. Parameters ---------- s : iterable (list or np.ndarray) Binary label (whether each element is labeled or not) in pu learning. prob_s_eq_1 : iterable (list or np.ndarray) The probability, for each example, whether it has label=1 P(label=1|x) Output (float) ------ Claesen's estimate for f1 in the pulearning setting.""" pred = np.asarray(prob_s_eq_1) >= 0.5 true_positives = sum((np.asarray(s) == 1) & (np.asarray(pred) == 1)) all_positives = sum(s) recall = true_positives / float(all_positives) frac_positive = sum(pred) / float(len(s)) return recall**2 / (2.0 * frac_positive) if frac_positive != 0 else np.nan
[docs]def confusion_matrix(true, pred) -> np.ndarray: """Implements a confusion matrix for true labels and predicted labels. true and pred MUST BE the same length and have the same distinct set of class labels represented. Results are identical (and similar computation time) to: "sklearn.metrics.confusion_matrix" However, this function avoids the dependency on sklearn. Parameters ---------- true : np.ndarray 1d Contains labels. Assumes true and pred contains the same set of distinct labels. pred : np.ndarray 1d A discrete vector of noisy labels, i.e. some labels may be erroneous. *Format requirements*: for dataset with K classes, labels must be in {0,1,...,K-1}. Returns ------- confusion_matrix : np.ndarray (2D) matrix of confusion counts with true on rows and pred on columns.""" assert len(true) == len(pred) true_classes = np.unique(true) pred_classes = np.unique(pred) K_true = len(true_classes) # Number of classes in true K_pred = len(pred_classes) # Number of classes in pred map_true = dict(zip(true_classes, range(K_true))) map_pred = dict(zip(pred_classes, range(K_pred))) result = np.zeros((K_true, K_pred)) for i in range(len(true)): result[map_true[true[i]]][map_pred[pred[i]]] += 1 return result
[docs]def compress_int_array(int_array, num_possible_values) -> np.ndarray: """Compresses dtype of np.ndarray<int> if num_possible_values is small enough.""" try: compressed_type = None if num_possible_values < np.iinfo(np.dtype("int16")).max: compressed_type = "int16" elif num_possible_values < np.iinfo(np.dtype("int32")).max: # pragma: no cover compressed_type = "int32" # pragma: no cover if compressed_type is not None: int_array = int_array.astype(compressed_type) return int_array except Exception: # int_array may not even be numpy array, keep as is then return int_array
[docs]def train_val_split( X, labels, train_idx, holdout_idx ) -> Tuple[DatasetLike, DatasetLike, LabelLike, LabelLike]: """Splits data into training/validation sets based on given indices""" labels_train, labels_holdout = ( labels[train_idx], labels[holdout_idx], ) # labels are always np.ndarray split_completed = False if isinstance(X, (pd.DataFrame, pd.Series)): X_train, X_holdout = X.iloc[train_idx], X.iloc[holdout_idx] split_completed = True if not split_completed: try: # check if X is pytorch Dataset object using lazy import import torch if isinstance(X, torch.utils.data.Dataset): # special splitting for pytorch Dataset X_train = torch.utils.data.Subset(X, train_idx) X_holdout = torch.utils.data.Subset(X, holdout_idx) split_completed = True except Exception: pass if not split_completed: try: # check if X is tensorflow Dataset object using lazy import import tensorflow if isinstance(X, tensorflow.data.Dataset): # special splitting for tensorflow Dataset X_train = extract_indices_tf(X, train_idx, allow_shuffle=True) X_holdout = extract_indices_tf(X, holdout_idx, allow_shuffle=False) split_completed = True except Exception: pass if not split_completed: try: X_train, X_holdout = X[train_idx], X[holdout_idx] except Exception: raise ValueError( "Cleanlab cannot split this form of dataset (required for cross-validation). " "Try a different data format, " "or implement the cross-validation yourself and instead provide out-of-sample `pred_probs`" ) return X_train, X_holdout, labels_train, labels_holdout
[docs]def subset_X_y(X, labels, mask) -> Tuple[DatasetLike, LabelLike]: """Extracts subset of features/labels where mask is True""" labels = subset_labels(labels, mask) X = subset_data(X, mask) return X, labels
[docs]def subset_labels(labels, mask) -> Union[list, np.ndarray, pd.Series]: """Extracts subset of labels where mask is True""" try: # filtering labels as if it is array or DataFrame return labels[mask] except Exception: try: # filtering labels as if it is list return [l for idx, l in enumerate(labels) if mask[idx]] except Exception: raise TypeError("labels must be 1D np.ndarray, list, or pd.Series.")
[docs]def subset_data(X, mask) -> DatasetLike: """Extracts subset of data examples where mask (np.ndarray) is True""" try: import torch if isinstance(X, torch.utils.data.Dataset): mask_idx_list = list(np.nonzero(mask)[0]) return torch.utils.data.Subset(X, mask_idx_list) except Exception: pass try: with warnings.catch_warnings(): warnings.filterwarnings("ignore") import tensorflow if isinstance(X, tensorflow.data.Dataset): # special splitting for tensorflow Dataset mask_idx = np.nonzero(mask)[0] return extract_indices_tf(X, mask_idx, allow_shuffle=True) except Exception: pass try: return X[mask] except Exception: raise TypeError("Data features X must be subsettable with boolean mask array: X[mask]")
[docs]def extract_indices_tf(X, idx, allow_shuffle) -> DatasetLike: """Extracts subset of tensorflow dataset corresponding to examples at particular indices. Args: X : ``tensorflow.data.Dataset`` idx : array_like of integer indices corresponding to examples to keep in the dataset. Returns subset of examples in the dataset X that correspond to these indices. allow_shuffle : bool Whether or not shuffling of this data is allowed (eg. must turn off shuffling for validation data). Note: this code only works on Datasets in which: * ``shuffle()`` has been called before ``batch()``, * no other order-destroying operation (eg. ``repeat()``) has been applied. Indices are extracted from the original version of Dataset (before shuffle was called rather than in shuffled order). """ import tensorflow idx = np.asarray(idx) idx = np.int64(idx) # needed for Windows (reconsider if necessary in the future) og_batch_size = None if hasattr(X, "_batch_size"): og_batch_size = int(X._batch_size) X = X.unbatch() unshuffled_X, buffer_size = unshuffle_tensorflow_dataset(X) if unshuffled_X is not None: X = unshuffled_X # Create index,value pairs in the dataset (adds extra indices that werent there before) X = X.enumerate() keys_tensor = tensorflow.constant(idx) vals_tensor = tensorflow.ones_like(keys_tensor) # Ones will be casted to True table = tensorflow.lookup.StaticHashTable( tensorflow.lookup.KeyValueTensorInitializer(keys_tensor, vals_tensor), default_value=0, ) # If index not in table, return 0 def hash_table_filter(index, value): table_value = table.lookup(index) # 1 if index in arr, else 0 index_in_arr = tensorflow.cast(table_value, tensorflow.bool) # 1 -> True, 0 -> False return index_in_arr # Filter the dataset, then drop the added indices X_subset = X.filter(hash_table_filter).map(lambda idx, value: value) if (unshuffled_X is not None) and allow_shuffle: X_subset = X_subset.shuffle(buffer_size=buffer_size) if og_batch_size is not None: # reset batch size to original value X_subset = X_subset.batch(og_batch_size) return X_subset
[docs]def unshuffle_tensorflow_dataset(X) -> tuple: """Applies iterative inverse transformations to dataset to get version before ShuffleDataset was created. If no ShuffleDataset is in the transformation-history of this dataset, returns None. Parameters ---------- X : a tensorflow Dataset that may have been created via series of transformations, one being shuffle. Returns ------- Tuple (pre_X, buffer_size) where: pre_X : Dataset that was previously transformed to get ShuffleDataset (or None), buffer_size : int `buffer_size` previously used in ShuffleDataset, or ``len(pre_X)`` if buffer_size cannot be determined, or None if no ShuffleDataset found. """ try: from tensorflow.python.data.ops.dataset_ops import ( ShuffleDataset, ) X_inputs = [X] while len(X_inputs) == 1: pre_X = X_inputs[0] if isinstance(pre_X, ShuffleDataset): buffer_size = len(pre_X) if hasattr(pre_X, "_buffer_size"): buffer_size = pre_X._buffer_size.numpy() X_inputs = ( pre_X._inputs() ) # get the dataset that was transformed to create the ShuffleDataset if len(X_inputs) == 1: return (X_inputs[0], buffer_size) X_inputs = pre_X._inputs() # returns list of input datasets used to create X except Exception: pass return (None, None)
[docs]def is_torch_dataset(X) -> bool: try: import torch if isinstance(X, torch.utils.data.Dataset): return True except Exception: pass return False # assumes this cannot be torch dataset if torch cannot be imported
[docs]def is_tensorflow_dataset(X) -> bool: try: import tensorflow if isinstance(X, tensorflow.data.Dataset): return True except Exception: pass return False # assumes this cannot be tensorflow dataset if tensorflow cannot be imported
[docs]def csr_vstack(a, b) -> DatasetLike: """Takes in 2 csr_matrices and appends the second one to the bottom of the first one. Alternative to scipy.sparse.vstack. Returns a sparse matrix. """ a.data = np.hstack((a.data, b.data)) a.indices = np.hstack((a.indices, b.indices)) a.indptr = np.hstack((a.indptr, (b.indptr + a.nnz)[1:])) a._shape = (a.shape[0] + b.shape[0], b.shape[1]) return a
[docs]def append_extra_datapoint(to_data, from_data, index) -> DatasetLike: """Appends an extra datapoint to the data object ``to_data``. This datapoint is taken from the data object ``from_data`` at the corresponding index. One place this could be useful is ensuring no missing classes after train/validation split. """ if not (type(from_data) is type(to_data)): raise ValueError("Cannot append datapoint from different type of data object.") if isinstance(to_data, np.ndarray): return np.vstack([to_data, from_data[index]]) elif isinstance(from_data, (pd.DataFrame, pd.Series)): X_extra = from_data.iloc[[index]] # type: ignore to_data = pd.concat([to_data, X_extra]) return to_data.reset_index(drop=True) else: try: X_extra = from_data[index] try: return to_data.append(X_extra) except Exception: # special append for sparse matrix return csr_vstack(to_data, X_extra) except Exception: raise TypeError("Data features X must support: X.append(X[i])")
[docs]def get_num_classes(labels=None, pred_probs=None, label_matrix=None, multi_label=None) -> int: """Determines the number of classes based on information considered in a canonical ordering. label_matrix can be: noise_matrix, inverse_noise_matrix, confident_joint, or any other K x K matrix where K = number of classes. """ if pred_probs is not None: # pred_probs is number 1 source of truth return pred_probs.shape[1] if label_matrix is not None: # matrix dimension is number 2 source of truth if label_matrix.shape[0] != label_matrix.shape[1]: raise ValueError(f"label matrix must be K x K, not {label_matrix.shape}") else: return label_matrix.shape[0] if labels is None: raise ValueError("Cannot determine number of classes from None input") return num_unique_classes(labels, multi_label=multi_label)
[docs]def num_unique_classes(labels, multi_label=None) -> int: """Finds the number of unique classes for both single-labeled and multi-labeled labels. If multi_label is set to None (default) this method will infer if multi_label is True or False based on the format of labels. This allows for a more general form of multiclass labels that looks like this: [1, [1,2], [0], [0, 1], 2, 1]""" return len(get_unique_classes(labels, multi_label))
[docs]def get_unique_classes(labels, multi_label=None) -> set: """Returns the set of unique classes for both single-labeled and multi-labeled labels. If multi_label is set to None (default) this method will infer if multi_label is True or False based on the format of labels. This allows for a more general form of multiclass labels that looks like this: [1, [1,2], [0], [0, 1], 2, 1]""" if multi_label is None: multi_label = any(isinstance(l, list) for l in labels) if multi_label: return set(l for grp in labels for l in list(grp)) else: return set(labels)
[docs]def format_labels(labels: LabelLike) -> Tuple[np.ndarray, dict]: """Takes an array of labels and formats it such that labels are in the set ``0, 1, ..., K-1``, where ``K`` is the number of classes. The labels are assigned based on lexicographic order. This is useful for mapping string class labels to the integer format required by many cleanlab (and sklearn) functions. Returns ------- formatted_labels Returns np.ndarray of shape ``(N,)``. The return labels will be properly formatted and can be passed to other cleanlab functions. mapping A dictionary showing the mapping of new to old labels, such that ``mapping[k]`` returns the name of the k-th class. """ labels = labels_to_array(labels) if labels.ndim != 1: raise ValueError("labels must be 1D numpy array.") unique_labels = np.unique(labels) label_map = {label: i for i, label in enumerate(unique_labels)} formatted_labels = np.array([label_map[l] for l in labels]) inverse_map = {i: label for label, i in label_map.items()} return formatted_labels, inverse_map
[docs]def smart_display_dataframe(df): # pragma: no cover """Display a pandas dataframe if in a jupyter notebook, otherwise print it to console.""" try: from IPython.display import display display(df) except Exception: print(df)