Keras weighted categorical_crossentropy (please read comments for updated version)

keras_weighted_categorical_crossentropy.py

"""
A weighted version of categorical_crossentropy for keras (2.0.6). This lets you apply a weight to unbalanced classes.
@url: https://gist.github.com/wassname/ce364fddfc8a025bfab4348cf5de852d
@author: wassname
"""
from keras import backend as K
def weighted_categorical_crossentropy(weights):
    """
    A weighted version of keras.objectives.categorical_crossentropy
    
    Variables:
        weights: numpy array of shape (C,) where C is the number of classes
    
    Usage:
        weights = np.array([0.5,2,10]) # Class one at 0.5, class 2 twice the normal weights, class 3 10x.
        loss = weighted_categorical_crossentropy(weights)
        model.compile(loss=loss,optimizer='adam')
    """
    
    weights = K.variable(weights)
        
    def loss(y_true, y_pred):
        # scale predictions so that the class probas of each sample sum to 1
        y_pred /= K.sum(y_pred, axis=-1, keepdims=True)
        # clip to prevent NaN's and Inf's
        y_pred = K.clip(y_pred, K.epsilon(), 1 - K.epsilon())
        # calc
        loss = y_true * K.log(y_pred) * weights
        loss = -K.sum(loss, -1)
        return loss
    
    return loss

import numpy as np
from keras.activations import softmax
from keras.objectives import categorical_crossentropy

# init tests
samples=3
maxlen=4
vocab=5

y_pred_n = np.random.random((samples,maxlen,vocab)).astype(K.floatx())
y_pred = K.variable(y_pred_n)
y_pred = softmax(y_pred)

y_true_n = np.random.random((samples,maxlen,vocab)).astype(K.floatx())
y_true = K.variable(y_true_n)
y_true = softmax(y_true)

# test 1 that it works the same as categorical_crossentropy with weights of one
weights = np.ones(vocab)

loss_weighted=weighted_categorical_crossentropy(weights)(y_true,y_pred).eval(session=K.get_session())
loss=categorical_crossentropy(y_true,y_pred).eval(session=K.get_session())
np.testing.assert_almost_equal(loss_weighted,loss)
print('OK test1')


# test 2 that it works differen't than categorical_crossentropy with weights of less than one
weights = np.array([0.1,0.3,0.5,0.3,0.5])

loss_weighted=weighted_categorical_crossentropy(weights)(y_true,y_pred).eval(session=K.get_session())
loss=categorical_crossentropy(y_true,y_pred).eval(session=K.get_session())
np.testing.assert_array_less(loss_weighted,loss)
print('OK test2')

# same keras version as I tested it on?
import keras
assert keras.__version__.split('.')[:2]==['2', '0'], 'this was tested on keras 2.0.6 you have %s' % keras.__version
print('OK version')

test3.py

'''
test weighted_categorical_crossentropy on a real dataset
'''

from __future__ import print_function
import keras
from keras.datasets import cifar10
from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten
from keras.layers import Conv2D, MaxPooling2D

import os
import pickle
import numpy as np

batch_size = 32
num_classes = 10
epochs = 200
data_augmentation = False
num_predictions = 20
save_dir = os.path.join(os.getcwd(), 'saved_models')
model_name = 'keras_cifar10_trained_model.h5'

# The data, shuffled and split between train and test sets:
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')

# Convert class vectors to binary class matrices.
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)

model = Sequential()

model.add(Conv2D(32, (3, 3), padding='same',
                 input_shape=x_train.shape[1:]))
model.add(Activation('relu'))
model.add(Conv2D(32, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))

model.add(Conv2D(64, (3, 3), padding='same'))
model.add(Activation('relu'))
model.add(Conv2D(64, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))

model.add(Flatten())
model.add(Dense(512))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes))
model.add(Activation('softmax'))

# initiate RMSprop optimizer
opt = keras.optimizers.rmsprop(lr=0.0001, decay=1e-6)

# Let's train the model using RMSprop
weights = np.ones((10,))
model.compile(loss=weighted_categorical_crossentropy(weights),
              optimizer=opt,
              metrics=['accuracy'])

x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255

nc = 100
x_train = x_train[:nc]
y_train = y_train[:nc]
x_test = x_test[:nc]
y_test = y_test[:nc]

if not data_augmentation:
    print('Not using data augmentation.')
    model.fit(x_train, y_train,
              batch_size=batch_size,
              epochs=epochs,
              validation_data=(x_test, y_test),
              shuffle=True)
else:
    print('Using real-time data augmentation.')
    # This will do preprocessing and realtime data augmentation:
    datagen = ImageDataGenerator(
        featurewise_center=False,  # set input mean to 0 over the dataset
        samplewise_center=False,  # set each sample mean to 0
        featurewise_std_normalization=False,  # divide inputs by std of the dataset
        samplewise_std_normalization=False,  # divide each input by its std
        zca_whitening=False,  # apply ZCA whitening
        rotation_range=0,  # randomly rotate images in the range (degrees, 0 to 180)
        width_shift_range=0.1,  # randomly shift images horizontally (fraction of total width)
        height_shift_range=0.1,  # randomly shift images vertically (fraction of total height)
        horizontal_flip=True,  # randomly flip images
        vertical_flip=False)  # randomly flip images

    # Compute quantities required for feature-wise normalization
    # (std, mean, and principal components if ZCA whitening is applied).
    datagen.fit(x_train)

    # Fit the model on the batches generated by datagen.flow().
    model.fit_generator(datagen.flow(x_train, y_train,
                                     batch_size=batch_size),
                        steps_per_epoch=x_train.shape[0] // batch_size,
                        epochs=epochs,
                        validation_data=(x_test, y_test))

# Save model and weights
if not os.path.isdir(save_dir):
    os.makedirs(save_dir)
model_path = os.path.join(save_dir, model_name)
model.save(model_path)
print('Saved trained model at %s ' % model_path)

# Load label names to use in prediction results
label_list_path = 'datasets/cifar-10-batches-py/batches.meta'


keras_dir = os.path.expanduser(os.path.join('~', '.keras'))
datadir_base = os.path.expanduser(keras_dir)
if not os.access(datadir_base, os.W_OK):
    datadir_base = os.path.join('/tmp', '.keras')
label_list_path = os.path.join(datadir_base, label_list_path)

with open(label_list_path, mode='rb') as f:
    labels = pickle.load(f)

# Evaluate model with test data set and share sample prediction results
evaluation = model.evaluate_generator(datagen.flow(x_test, y_test,
                                      batch_size=batch_size),
                                      steps=x_test.shape[0] // batch_size)

print('Model Accuracy = %.2f' % (evaluation[1]))
nc=200
predict_gen = model.predict_generator(datagen.flow(x_test, y_test,
                                      batch_size=batch_size),
                                      steps=x_test.shape[0] // batch_size)

for predict_index, predicted_y in enumerate(predict_gen):
    actual_label = labels['label_names'][np.argmax(y_test[predict_index])]
    predicted_label = labels['label_names'][np.argmax(predicted_y)]
    print('Actual Label = %s vs. Predicted Label = %s' % (actual_label,
                                                          predicted_label))
    if predict_index == num_predictions:
        break
        
"""
Epoch 195/200
100/100 [==============================] - 2s - loss: 0.2921 - acc: 0.9300 - val_loss: 3.1197 - val_acc: 0.2300
Epoch 196/200
100/100 [==============================] - 2s - loss: 0.3474 - acc: 0.9300 - val_loss: 3.1419 - val_acc: 0.2200
Epoch 197/200
100/100 [==============================] - 2s - loss: 0.3614 - acc: 0.9000 - val_loss: 3.2418 - val_acc: 0.2300
Epoch 198/200
100/100 [==============================] - 2s - loss: 0.4221 - acc: 0.8800 - val_loss: 3.1150 - val_acc: 0.2100
Epoch 199/200
100/100 [==============================] - 2s - loss: 0.3901 - acc: 0.8900 - val_loss: 3.1687 - val_acc: 0.2400
Epoch 200/200
100/100 [==============================] - 2s - loss: 0.3228 - acc: 0.9400 - val_loss: 3.3791 - val_acc: 0.2200
Saved trained model at D:\NotBackedUp\MyDocumentsLarge_mclark52\WinPython-64bit-3.5.3.1Qt5\notebooks\saved_models\keras_cifar10_trained_model.h5 
Model Accuracy = 0.21
Actual Label = cat vs. Predicted Label = ship
Actual Label = ship vs. Predicted Label = cat
Actual Label = ship vs. Predicted Label = truck
Actual Label = airplane vs. Predicted Label = dog
Actual Label = frog vs. Predicted Label = bird
Actual Label = frog vs. Predicted Label = horse
Actual Label = automobile vs. Predicted Label = truck
Actual Label = frog vs. Predicted Label = airplane
Actual Label = cat vs. Predicted Label = automobile
Actual Label = automobile vs. Predicted Label = horse
Actual Label = airplane vs. Predicted Label = airplane
Actual Label = truck vs. Predicted Label = truck
Actual Label = dog vs. Predicted Label = bird
Actual Label = horse vs. Predicted Label = truck
Actual Label = truck vs. Predicted Label = bird
Actual Label = ship vs. Predicted Label = truck
Actual Label = dog vs. Predicted Label = truck
Actual Label = horse vs. Predicted Label = bird
Actual Label = ship vs. Predicted Label = automobile
Actual Label = frog vs. Predicted Label = cat
Actual Label = horse vs. Predicted Label = automobile
"""

Can confirm that this gist works correctly in TensorFlow 2.x + Keras in the task of multiclass segmentation. You have to assign the weights in range [0..1]. I've calculated the ratios for each class in the dataset (by pixels), the less frequent class got weight of 1, all the others got 0.xx

I had to add tf.cast(y_true, dtype=tf.float32) for it to work, other than that thank you very much! Would you happen to have one for binary cross-entropy (in case of multi-label classification it would be useful)

Somehow I think the weights should be applied on the logits rather than the probabilities so that the final sum of all probabilities are still 1.

I have a basic question, why the function take care of only the ones with this?
loss = y_true * K.log(y_pred) * weights
i think it's needed something like that:
loss = y_true * K.log(y_pred) * weights
loss += (1 - y_true) * K.log(1 - y_pred) * weights

This does not work for me. I get this error:
InvalidArgumentError: cannot compute Mul as input #1(zero-based) was expected to be a uint8 tensor but is a float tensor [Op:Mul]

Does anyone know how to fix this? I see angeligareta says to use tf.cast(y_true, dtype=tf.float32), but I still get the same error.

For me, I had to replace weights = K.variable( weights ) to weights = K.constant( weights ).

Thanks for this! I'm curious, is there any difference with the categorical_focal_crossentropy function?