Implementation of the AdamW optimizer(Ilya Loshchilov, Frank Hutter) for Keras.

• python 3.6
• Keras 2.1.6
• tensorflow(-gpu) 1.8.0

Additionally to a usual Keras setup for neural nets building (see Keras for details)

from AdamW import AdamW

adamw = AdamW(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0., weight_decay=0.025, batch_size=1, samples_per_epoch=1, epochs=1)

Then nothing change compared to the usual usage of an optimizer in Keras after the definition of a model's architecture

model = Sequential()
<definition of the model_architecture>
model.compile(loss="mse", optimizer=adamw, metrics=[metrics.mse], ...)

Note that the size of a batch (batch_size), number of training samples per epoch (samples_per_epoch) and the number of epochs (epochs) are necessary to the normalization of the weight decay (paper, Section 4)

• Weight decay added to the parameters optimization
• Normalized weight decay added

• Cosine annealing
• Warm restarts

ADAM: A METHOD FOR STOCHASTIC OPTIMIZATION, D.P. Kingma, J. Lei Ba

Fixing Weight Decay Regularization in Adam, I. Loshchilov, F. Hutter