tf.keras.losses.CategoricalFocalCrossentropy

Computes the alpha balanced focal crossentropy loss.

Inherits From: Loss

tf.keras.losses.CategoricalFocalCrossentropy(
    alpha=0.25,
    gamma=2.0,
    from_logits=False,
    label_smoothing=0.0,
    axis=-1,
    reduction=losses_utils.ReductionV2.AUTO,
    name='categorical_focal_crossentropy'
)

Use this crossentropy loss function when there are two or more label classes and if you want to handle class imbalance without using class_weights. We expect labels to be provided in a one_hot representation.

According to Lin et al., 2018, it helps to apply a focal factor to down-weight easy examples and focus more on hard examples. The general formula for the focal loss (FL) is as follows:

FL(p_t) = (1 − p_t)^gamma * log(p_t)

where p_t is defined as follows: p_t = output if y_true == 1, else 1 - output

(1 − p_t)^gamma is the modulating_factor, where gamma is a focusing parameter. When gamma = 0, there is no focal effect on the cross entropy. gamma reduces the importance given to simple examples in a smooth manner.

The authors use alpha-balanced variant of focal loss (FL) in the paper: FL(p_t) = −alpha * (1 − p_t)^gamma * log(p_t)

where alpha is the weight factor for the classes. If alpha = 1, the loss won't be able to handle class imbalance properly as all classes will have the same weight. This can be a constant or a list of constants. If alpha is a list, it must have the same length as the number of classes.

The formula above can be generalized to: FL(p_t) = alpha * (1 − p_t)^gamma * CrossEntropy(y_true, y_pred)

where minus comes from CrossEntropy(y_true, y_pred) (CE).

Extending this to multi-class case is straightforward: FL(p_t) = alpha * (1 − p_t)^gamma * CategoricalCE(y_true, y_pred)

In the snippet below, there is # classes floating pointing values per example. The shape of both y_pred and y_true are [batch_size, num_classes].

Standalone usage:

y_true = [[0., 1., 0.], [0., 0., 1.]]
y_pred = [[0.05, 0.95, 0], [0.1, 0.8, 0.1]]
# Using 'auto'/'sum_over_batch_size' reduction type.
cce = tf.keras.losses.CategoricalFocalCrossentropy()
cce(y_true, y_pred).numpy()
0.23315276
 
# Calling with 'sample_weight'.
cce(y_true, y_pred, sample_weight=tf.constant([0.3, 0.7])).numpy()
0.1632
 
# Using 'sum' reduction type.
cce = tf.keras.losses.CategoricalFocalCrossentropy(
    reduction=tf.keras.losses.Reduction.SUM)
cce(y_true, y_pred).numpy()
0.46631
 
# Using 'none' reduction type.
cce = tf.keras.losses.CategoricalFocalCrossentropy(
    reduction=tf.keras.losses.Reduction.NONE)
cce(y_true, y_pred).numpy()
array([3.2058331e-05, 4.6627346e-01], dtype=float32)

Usage with the compile() API:

model.compile(optimizer='adam',
              loss=tf.keras.losses.CategoricalFocalCrossentropy())

alpha A weight balancing factor for all classes, default is 0.25 as mentioned in the reference. It can be a list of floats or a scalar. In the multi-class case, alpha may be set by inverse class frequency by using compute_class_weight from sklearn.utils.
gamma A focusing parameter, default is 2.0 as mentioned in the reference. It helps to gradually reduce the importance given to simple (easy) examples in a smooth manner.
from_logits Whether output is expected to be a logits tensor. By default, we consider that output encodes a probability distribution.
label_smoothing Float in [0, 1]. When > 0, label values are smoothed, meaning the confidence on label values are relaxed. For example, if 0.1, use 0.1 / num_classes for non-target labels and 0.9 + 0.1 / num_classes for target labels.
axis The axis along which to compute crossentropy (the features axis). Defaults to -1.
reduction Type of tf.keras.losses.Reduction to apply to loss. Default value is AUTO. AUTO indicates that the reduction option will be determined by the usage context. For almost all cases this defaults to SUM_OVER_BATCH_SIZE. When used under a tf.distribute.Strategy, except via Model.compile() and Model.fit(), using AUTO or SUM_OVER_BATCH_SIZE will raise an error. Please see this custom training tutorial for more details.
name Optional name for the instance. Defaults to 'categorical_focal_crossentropy'.

Methods

from_config

View source

@classmethod
from_config(
    config
)

Instantiates a Loss from its config (output of get_config()).

Args
config Output of get_config().

Returns
A keras.losses.Loss instance.

get_config

View source

get_config()

Returns the config dictionary for a Loss instance.

__call__

View source

__call__(
    y_true, y_pred, sample_weight=None
)

Invokes the Loss instance.

Args
y_true Ground truth values. shape = [batch_size, d0, .. dN], except sparse loss functions such as sparse categorical crossentropy where shape = [batch_size, d0, .. dN-1]
y_pred The predicted values. shape = [batch_size, d0, .. dN]
sample_weight Optional sample_weight acts as a coefficient for the loss. If a scalar is provided, then the loss is simply scaled by the given value. If sample_weight is a tensor of size [batch_size], then the total loss for each sample of the batch is rescaled by the corresponding element in the sample_weight vector. If the shape of sample_weight is [batch_size, d0, .. dN-1] (or can be broadcasted to this shape), then each loss element of y_pred is scaled by the corresponding value of sample_weight. (Note ondN-1: all loss functions reduce by 1 dimension, usually axis=-1.)

Returns
Weighted loss float Tensor. If reduction is NONE, this has shape [batch_size, d0, .. dN-1]; otherwise, it is scalar. (Note dN-1 because all loss functions reduce by 1 dimension, usually axis=-1.)

Raises
ValueError If the shape of sample_weight is invalid.