tf.IndexedSlices gradients and dense gradients in keras

General Discussion
,
#1

Hi,
All existing optimizers in keras (adam, sgd, nadam, …) distinguish between tf.IndexedSlices gradients and dense gradients in order to update variables (in update_step() function):
def update_step(self, gradient, variable):
if isinstance(gradient, tf.IndexedSlices):
update manner 1
else:# Dense gradients.
update manner 2 (other update manner)

I want to write a custom optimizer and I want why should I distinguish between tf.IndexedSlices gradients and dense gradients?
thank you

#3

Hi @dali_dali, tf.IndexedSlices gradients are used when computing gradients for sparse tensors, whereas dense gradients are used when computing gradients for dense tensors. Distinguish between them allows you to use the appropriate gradient representation for the type of tensor being used. Thank You.

#5

Amazing information!