Skip to content
This repository has been archived by the owner on Aug 5, 2022. It is now read-only.

Caffe_v1.0.7
Compare
Choose a tag to compare
@hshen14 hshen14 released this 13 Dec 09:04
· 430 commits to master since this release
1.0.7
f6a2a6b
  • Features
  1. [Multi-node] Support weight gradient compression for better scaling efficiency of models with large FC layers, like VGG
  2. [Multi-node] Integrate LARS (Layer-wise Adaptive Rate Scaling) and apply to Alexnet with BatchNorm layers on 32K global batch size
  3. Enable pinning internal thread to cores for more stable training performance, e.g., data loader thread
  4. Merge pull request of supporting Flow LRCN from Github
  5. Support label smoothing regularization (idea from Inception-V3)
  • Bug fixings
  1. [Multi-node] Fix learning rate message and start first iteration from zero on multi-node to be consistent with single node
  2. Bug fixes on single node
  • Misc
  1. Upgrade MKLML to 2018.0.1.20171007 and MLSL to V2
  2. Enhance installation and benchmarking scripts
  3. Update the optimized models
Assets 3
Loading