Revision 5f613de6d19146e667dc200335b3a357119edf17 authored by KeDengMS on 21 September 2017, 06:08:38 UTC, committed by KeDengMS on 21 September 2017, 06:08:38 UTC
1. local learner update should happen before block sync 2. checkpoint save/restore block smoothed gradient 3. momentum should be calculated on actual number of samples
1 parent 23dbaad
File | Mode | Size |
---|---|---|
.gitattributes | -rw-r--r-- | 39 bytes |
AllReduceDistGradAggregator.h | -rw-r--r-- | 29.9 KB |
BlockMomentumDistributedLearner.h | -rw-r--r-- | 21.6 KB |
BlockMomentumSGD.h | -rw-r--r-- | 14.2 KB |
LICENSE-GENERAL.md | -rw-r--r-- | 7.6 KB |
LICENSE-NON-COMMERCIAL.md | -rw-r--r-- | 6.1 KB |
LICENSE-README.md | -rw-r--r-- | 539 bytes |
MatrixQuantizer.h | -rw-r--r-- | 2.6 KB |
QuantizedDataParallelDistributedLearner.h | -rw-r--r-- | 4.2 KB |
QuantizedDistributedCommunicator.h | -rw-r--r-- | 29.2 KB |
README.md | -rw-r--r-- | 668 bytes |
V2AllReduceDistGradAggregator.h | -rw-r--r-- | 13.0 KB |
V2BlockMomentumSGD.h | -rw-r--r-- | 16.6 KB |
Computing file changes ...