tff.learning.optimizers.handle_indexed_slices_gradients

Converts any tf.IndexedSlices to tensors.

tff.learning.optimizers.handle_indexed_slices_gradients(
    gradients: _Structure[T]
) -> _Structure[Union[T, tf.IndexedSlices]]

The tf.IndexedSlices class is used principally in the definition of gradients for operations that have sparse gradients (e.g. tf.gather). See also tf.GradientTape documentation. This method is an elementary utility converting the slices to a tensor, which can be used to make an optimizer immediately compatible with such gradients. All other values are left unmodified.

Note however, this operation may be expensive in some situations. For more details, see https://github.com/tensorflow/tensorflow/blob/2b44549aca184ae0eb986a8bd46feef2b17004ab/tensorflow/python/framework/indexed_slices.py#L406

gradients A collection of gradients to be used by an optimizer.

The same collection with tf.IndexedSlices replaced by tensors.