wbcefn.py

@tf.keras.saving.register_keras_serializable(name="weighted_binary_crossentropy")
def weighted_binary_crossentropy(target, output, weights):
    target = tf.convert_to_tensor(target)
    output = tf.convert_to_tensor(output)
    weights = tf.convert_to_tensor(weights, dtype=target.dtype)

    epsilon_ = tf.constant(tf.keras.backend.epsilon(), output.dtype.base_dtype)
    output = tf.clip_by_value(output, epsilon_, 1.0 - epsilon_)

    # Compute cross entropy from probabilities.
    bce = weights[1] * target * tf.math.log(output + epsilon_)
    bce += weights[0] * (1 - target) * tf.math.log(1 - output + epsilon_)
    return -bce