MCPcopy Create free account
hub / github.com/apache/singa / sparsification

Method sparsification

python/singa/opt.py:786–801  ·  view source on GitHub ↗

Performs all reduce of a tensor after sparsification. Args: tensor(Tensor): a tensor to be all-reduced accumulation(Tensor): local gradient accumulation spars(float): a parameter to control sparsity as defined below topK(bool):

(self, tensor, accumulation, spars, topK)

Source from the content-addressed store, hash-verified

784 self.communicator.fusedSynchHalf(tensor, send)
785
786 def sparsification(self, tensor, accumulation, spars, topK):
787 """Performs all reduce of a tensor after sparsification.
788
789 Args:
790 tensor(Tensor): a tensor to be all-reduced
791 accumulation(Tensor): local gradient accumulation
792 spars(float): a parameter to control sparsity as defined below
793 topK(bool): When topK is False, it sparsifies the gradient with absolute
794 value >= sparsWhen topK is True, it sparsifies a fraction of total gradient
795 number equals to spars, E.g. when spars = 0.01, it sparsifies 1 % of the
796 total gradient elements
797 """
798 if accumulation is None:
799 self.communicator.sparsification(tensor, spars, topK)
800 else:
801 self.communicator.sparsification(tensor, accumulation, spars, topK)
802
803 def fused_sparsification(self, tensor, accumulation, spars, topK):
804 """Performs all reduce of the tensors after fusing and sparsification.

Callers 1

Calls

no outgoing calls

Tested by

no test coverage detected