tf.keras.losses.BinaryFocalCrossentropy

Computes focal cross-entropy loss between true labels and predictions.

Inherits From: Loss

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

Binary cross-entropy loss is often used for binary (0 or 1) classification tasks. The loss function requires the following inputs:

  • y_true (true label): This is either 0 or 1.
  • y_pred (predicted value): This is the model's prediction, i.e, a single floating-point value which either represents a logit, (i.e, value in [-inf, inf] when from_logits=True) or a probability (i.e, value in [0., 1.] when from_logits=False).

According to Lin et al., 2018, it helps to apply a "focal factor" to down-weight easy examples and focus more on hard examples. By default, the focal tensor is computed as follows:

focal_factor = (1 - output) ** gamma for class 1 focal_factor = output ** gamma for class 0 where gamma is a focusing parameter. When gamma=0, this function is equivalent to the binary crossentropy loss.

With the compile() API:

model.compile(
  loss=tf.keras.losses.BinaryFocalCrossentropy(gamma=2.0, from_logits=True),
  ....
)

As a standalone function:

# Example 1: (batch_size = 1, number of samples = 4)
y_true = [0, 1, 0, 0]
y_pred = [-18.6, 0.51, 2.94, -12.8]
loss = tf.keras.losses.BinaryFocalCrossentropy(gamma=2,
                                               from_logits=True)
loss(y_true, y_pred).numpy()
0.691

# Apply class weight
loss = tf.keras.losses.BinaryFocalCrossentropy(
    apply_class_balancing=True, gamma=2, from_logits=True)
loss(y_true, y_pred).numpy()
0.51

# Example 2: (batch_size = 2, number of samples = 4)
y_true = [[0, 1], [0, 0]]
y_pred = [[-18.6, 0.51], [2.94, -12.8]]
# Using default 'auto'/'sum_over_batch_size' reduction type.
loss = tf.keras.losses.BinaryFocalCrossentropy(gamma=3,
                                               from_logits=True)
loss(y_true, y_pred).numpy()
0.647

# Apply class weight
loss = tf.keras.losses.BinaryFocalCrossentropy(
    apply_class_balancing=True, gamma=3, from_logits=True)
loss(y_true, y_pred).numpy()
0.482

# Using 'sample_weight' attribute with focal effect
loss = tf.keras.losses.BinaryFocalCrossentropy(gamma=3,
                                               from_logits=True)
loss(y_true, y_pred, sample_weight=[0.8, 0.2]).numpy()
0.133

# Apply class weight
loss = tf.keras.losses.BinaryFocalCrossentropy(
    apply_class_balancing=True, gamma=3, from_logits=True)
loss(y_true, y_pred, sample_weight=[0.8, 0.2]).numpy()
0.097

# Using 'sum' reduction` type.
loss = tf.keras.losses.BinaryFocalCrossentropy(gamma=4,
                                               from_logits=True,
    reduction=tf.keras.losses.Reduction.SUM)
loss(y_true, y_pred).numpy()
1.222

# Apply class weight
loss = tf.keras.losses.BinaryFocalCrossentropy(
    apply_class_balancing=True, gamma=4, from_logits=True,
    reduction=tf.keras.losses.Reduction.SUM)
loss(y_true, y_pred).numpy()
0.914

# Using 'none' reduction type.
loss = tf.keras.losses.BinaryFocalCrossentropy(
    gamma=5, from_logits=True,
    reduction=tf.keras.losses.Reduction.NONE)
loss(y_true, y_pred).numpy()
array([0.0017 1.1561], dtype=float32)

# Apply class weight
loss = tf.keras.losses.BinaryFocalCrossentropy(
    apply_class_balancing=True, gamma=5, from_logits=True,
    reduction=tf.keras.losses.Reduction.NONE)
loss(y_true, y_pred).numpy()
array([0.0004 0.8670], dtype=float32)

apply_class_balancing A bool, whether to apply weight balancing on the binary classes 0 and 1.
alpha A weight balancing factor for class 1, default is 0.25 as mentioned in reference Lin et al., 2018. The weight for class 0 is 1.0 - alpha.
gamma A focusing parameter used to compute the focal factor, default is 2.0 as mentioned in the reference Lin et al., 2018.
from_logits Whether to interpret y_pred as a tensor of logit values. By default, we assume that y_pred are probabilities (i.e., values in [0, 1]).
label_smoothing Float in [0, 1]. When 0, no smoothing occurs. When > 0, we compute the loss between the predicted labels and a smoothed version of the true labels, where the smoothing squeezes the labels towards 0.5. Larger values of label_smoothing correspond to heavier smoothing.
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 Name for the op. Defaults to 'binary_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.