tfr.keras.losses.GumbelApproxNDCGLoss

Computes the Gumbel approximate NDCG loss between y_true and y_pred.

Inherits From: ApproxNDCGLoss

Implementation of Gumbel ApproxNDCG loss (Bruch et al, 2020). This loss is the same as tfr.keras.losses.ApproxNDCGLoss but where logits are sampled from the Gumbel distribution:

y_new_pred ~ Gumbel(y_pred, 1 / temperature)

Standalone usage:

tf.random.set_seed(42)
y_true = [[1., 0.]]
y_pred = [[0.6, 0.8]]
loss = tfr.keras.losses.GumbelApproxNDCGLoss(seed=42)
loss(y_true, y_pred).numpy()
-0.8160851
# Using a higher gumbel temperature
loss = tfr.keras.losses.GumbelApproxNDCGLoss(gumbel_temperature=2.0,
    seed=42)
loss(y_true, y_pred).numpy()
-0.7583889
# 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.GumbelApproxNDCGLoss(seed=42, ragged=True)
loss(y_true, y_pred).numpy()
-0.6987189

Usage with the compile() API:

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

Definition:

\[\mathcal{L}(\{y\}, \{s\}) = \text{ApproxNDCGLoss}(\{y\}, \{z\})\]

where

\[ z \sim \text{Gumbel}(s, \beta)\\ p(z) = e^{-t-e^{-t} }\\ t = \beta(z - s)\\ \beta = \frac{1}{\text{temperature} } \]

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.