adashift icon indicating copy to clipboard operation
adashift copied to clipboard

AdaShift optimizer implementation in PyTorch

Adashift

Reproducing of the AdaShift: Decorrelation and Convergence of Adaptive Learning Rate Methods as a part of ICLR Reproducibility Challange 2019. See our report.

Experiments

Synthetic Experiment

Synthetic Synthetic_optval

Logistic Regression on MNIST

LR1 LR2

W-GAN

wgan-loss wgan-inception-score

fixed-gen

NMT

nmt

Dependencies

  • Python 3.4+,
  • NumPy, Pandas
  • PyTorch 0.4+
  • SciPy, Matplotlib
  • OpenNMT-py for NMT
  • GPU for DL experiments

References

On the Convergence of Adam and Beyond

Adam

AdaShift: Decorrelation and Convergence of Adaptive Learning Rate Methods