Scales per-example losses with sample_weights and computes their average.
tf.nn.compute_average_loss( per_example_loss, sample_weight=None, global_batch_size=None )
Usage with distribution strategy and custom training loop:
with strategy.scope(): def compute_loss(labels, predictions, sample_weight=None): # If you are using a `Loss` class instead, set reduction to `NONE` so that # we can do the reduction afterwards and divide by global batch size. per_example_loss = tf.keras.losses.sparse_categorical_crossentropy( labels, predictions) # Compute loss that is scaled by sample_weight and by global batch size. return tf.compute_average_loss( per_example_loss, sample_weight=sample_weight, global_batch_size=GLOBAL_BATCH_SIZE)
per_example_loss: Per-example loss.
sample_weight: Optional weighting for each example.
global_batch_size: Optional global batch size value. Defaults to (size of first dimension of
losses) * (number of replicas).
Scalar loss value.