Method to generate class weights given a multi-class or multi-label set of classes using Python, supporting one-hot-encoded labels.

generate_class_weights.py

import numpy as np
from sklearn.utils.class_weight import compute_class_weight
from sklearn.preprocessing import MultiLabelBinarizer


def generate_class_weights(class_series, multi_class=True, one_hot_encoded=False):
  """
  Method to generate class weights given a set of multi-class or multi-label labels, both one-hot-encoded or not.

  Some examples of different formats of class_series and their outputs are:
    - generate_class_weights(['mango', 'lemon', 'banana', 'mango'], multi_class=True, one_hot_encoded=False)
    {'banana': 1.3333333333333333, 'lemon': 1.3333333333333333, 'mango': 0.6666666666666666}
    - generate_class_weights([[1, 0, 0], [0, 1, 0], [0, 0, 1], [1, 0, 0]], multi_class=True, one_hot_encoded=True)
    {0: 0.6666666666666666, 1: 1.3333333333333333, 2: 1.3333333333333333}
    - generate_class_weights([['mango', 'lemon'], ['mango'], ['lemon', 'banana'], ['lemon']], multi_class=False, one_hot_encoded=False)
    {'banana': 1.3333333333333333, 'lemon': 0.4444444444444444, 'mango': 0.6666666666666666}
    - generate_class_weights([[0, 1, 1], [0, 0, 1], [1, 1, 0], [0, 1, 0]], multi_class=False, one_hot_encoded=True)
    {0: 1.3333333333333333, 1: 0.4444444444444444, 2: 0.6666666666666666}

  The output is a dictionary in the format { class_label: class_weight }. In case the input is one hot encoded, the class_label would be index
  of appareance of the label when the dataset was processed. 
  In multi_class this is np.unique(class_series) and in multi-label np.unique(np.concatenate(class_series)).

  Author: Angel Igareta ([email protected])
  """
  if multi_class:
    # If class is one hot encoded, transform to categorical labels to use compute_class_weight   
    if one_hot_encoded:
      class_series = np.argmax(class_series, axis=1)
  
    # Compute class weights with sklearn method
    class_labels = np.unique(class_series)
    class_weights = compute_class_weight(class_weight='balanced', classes=class_labels, y=class_series)
    return dict(zip(class_labels, class_weights))
  else:
    # It is neccessary that the multi-label values are one-hot encoded
    mlb = None
    if not one_hot_encoded:
      mlb = MultiLabelBinarizer()
      class_series = mlb.fit_transform(class_series)

    n_samples = len(class_series)
    n_classes = len(class_series[0])

    # Count each class frequency
    class_count = [0] * n_classes
    for classes in class_series:
        for index in range(n_classes):
            if classes[index] != 0:
                class_count[index] += 1
    
    # Compute class weights using balanced method
    class_weights = [n_samples / (n_classes * freq) if freq > 0 else 1 for freq in class_count]
    class_labels = range(len(class_weights)) if mlb is None else mlb.classes_
    return dict(zip(class_labels, class_weights))

Thank you. You can also use class_count = np.array(class_series).sum(axis=0) to count the labels in one line