Averaging Weights Leads to Wider Optima and Better Generalization

Code

• github.com/timgaripov/swa
  OfficialIn paperpytorch★ 0
• github.com/ENSTA-U2IS-AI/torch-uncertainty
  pytorch★ 484
• github.com/pycroscopy/atomai
  pytorch★ 225
• github.com/cmu-enyac/Renofeation
  pytorch★ 19
• github.com/dice-group/aswa
  pytorch★ 1
• github.com/julianfaraone/SWA
  pytorch★ 0
• github.com/izmailovpavel/torch_swa_examples
  pytorch★ 0
• github.com/benathi/fastswa-semi-sup
  pytorch★ 0
• github.com/wjmaddox/swa_gaussian
  pytorch★ 0
• github.com/Morxrc/kaggle_generate-dogs_public-1st-private-16th-solution
  none★ 0

Abstract

Deep neural networks are typically trained by optimizing a loss function with an SGD variant, in conjunction with a decaying learning rate, until convergence. We show that simple averaging of multiple points along the trajectory of SGD, with a cyclical or constant learning rate, leads to better generalization than conventional training. We also show that this Stochastic Weight Averaging (SWA) procedure finds much flatter solutions than SGD, and approximates the recent Fast Geometric Ensembling (FGE) approach with a single model. Using SWA we achieve notable improvement in test accuracy over conventional SGD training on a range of state-of-the-art residual networks, PyramidNets, DenseNets, and Shake-Shake networks on CIFAR-10, CIFAR-100, and ImageNet. In short, SWA is extremely easy to implement, improves generalization, and has almost no computational overhead.

Tasks

Benchmark Results

Reproductions