tfr.keras.losses.YetiLogisticLoss

Computes Yeti logistic loss between y_true and y_pred.

Adapted to neural network models from the Yeti loss implemenation for GBDT in (Lyzhin et al, 2022).

In this code base, we support Yeti loss with the DCG lambda weight option. The default uses the YetiDCGLambdaWeight with default settings. To customize, please set the lambda_weight to YetiDCGLambdaWeight.

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

loss = sum_a sum_i I[y_i > y_{i\pm 1}] * log(1 + exp(-(s^a_i - s^a_{i\pm 1})))

where

s^a_i = s_i + gumbel(0, 1)^a

Standalone usage:

y_true = [[1., 0.]]
y_pred = [[0.6, 0.8]]
loss = tfr.keras.losses.YetiLogisticLoss(sample_size=2, seed=1)
loss(y_true, y_pred).numpy()
0.90761846
# 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.YetiLogisticLoss(seed=1, ragged=True)
loss(y_true, y_pred).numpy()
0.43420443

Usage with the compile() API:

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

Definition:

\[ \mathcal{L}(\{y\}, \{s\}) = \sum_a \sum_i \sum_{j=i\pm 1}I[y_i > y_j] \log(1 + \exp(-(s^a_i - s^a_j))) \]

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 _RankingLoss.