def update_state(self, context: ClassifierContext) -> None:
"""
Update the internal state of the metric, using the information from the context object.
Args:
context (:py:class:`ashpy.contexts.ClassifierContext`): An AshPy Context holding
all the information the Metric needs.
"""
for features, labels in context.dataset:
predictions = context.classifier_model(
features, training=context.log_eval_mode == LogEvalMode.TRAIN)
self._distribute_strategy.experimental_run(
lambda: self._metric.update_state(
labels,
self._processing_predictions["fn"]
(predictions, **self._processing_predictions["kwargs"]),
))
def call(self, context: ClassifierContext, *, features: tf.Tensor,
labels: tf.Tensor, training: bool, **kwargs) -> tf.Tensor:
r"""
Compute the classifier loss.
Args:
context (:py:class:`ashpy.ClassifierContext`): Context for classification.
features (:py:class:`tf.Tensor`): Inputs for the classifier model.
labels (:py:class:`tf.Tensor`): Target for the classifier model.
training (bool): Whether is training or not.
**kwargs:
Returns:
:py:class:`tf.Tensor`: Loss value.
"""
predictions = context.classifier_model(features, training=training)
loss = self._fn(labels, predictions)
loss = tf.cond(
tf.equal(tf.rank(loss), tf.constant(4)),
lambda: loss,
lambda: tf.expand_dims(tf.expand_dims(loss, axis=-1), axis=-1),
)
return tf.reduce_mean(loss, axis=[1, 2])