tfr.keras.losses.UniqueSoftmaxLoss

Computes unique softmax cross-entropy loss between y_true and y_pred.

Implements unique rating softmax loss (Zhu et al, 2020).

For each list of scores s in y_pred and list of labels y in y_true:

loss = - sum_i (2^{y_i} - 1) *
               log(exp(s_i) / sum_j I(y_i > y_j) exp(s_j) + exp(s_i))

Standalone usage:

y_true = [[1., 0.]]
y_pred = [[0.6, 0.8]]
loss = tfr.keras.losses.UniqueSoftmaxLoss()
loss(y_true, y_pred).numpy()
0.7981389
# Using ragged tensors
y_true = tf.ragged.constant([[1., 0.], [0., 1., 0.]])
y_pred = tf.ragged.constant([[0.6, 0.8], [0.5, 0.8, 0.4]])
loss = tfr.keras.losses.UniqueSoftmaxLoss(ragged=True)
loss(y_true, y_pred).numpy()
0.83911896

Usage with the compile() API:

model.compile(optimizer='sgd', loss=tfr.keras.losses.UniqueSoftmaxLoss())

Definition:

\[ \mathcal{L}(\{y\}, \{s\}) = - \sum_i (2^{y_i} - 1) \log\left(\frac{\exp(s_i)}{\sum_j I_{y_i > y_j} \exp(s_j) + \exp(s_i)}\right) \]

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.

Methods

from_config

View source

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

Args
config Output of get_config().

Returns
A Loss instance.

get_config

View source

Returns the config dictionary for a Loss instance.

__call__

View source

See tf.keras.losses.Loss.