scape1989 / amsgrad-tensorflow

Simple Tensorflow implementation of "On the Convergence of Adam and Beyond" (ICLR 2018)

AMSGrad-Tensorflow

Simple Tensorflow implementation of On the Convergence of Adam and Beyond

Hyperparameter

• For the default hyperparameter, we set it to the best value in the experiment
• learning_rate = 0.01
• beta1 = 0.9
• beta2 = 0.99
• Depending on which network you are using, performance may be good at beta2 = 0.99 (default)

Usage

  from AMSGrad import AMSGrad
  
  train_op = AMSGrad(learning_rate=0.01, beta1=0.9, beta2=0.99, epsilon=1e-8).minimize(loss)

Network Architecture

  x = fully_connected(inputs=images, units=100)
  x = relu(x)
  logits = fully_connected(inputs=x, units=10)

Mnist Result (iteration = 30K)

lr=0.1, beta1=0.9, beta2=various

 

lr=0.01, beta1=0.9, beta2=various

 

Reference

• Keras-AMSGrad

Author

Junho Kim