This repository contains code for Cosine Annealing Scheduler based on SGDR: Stochastic Gradient Descent with Warm Restarts implemented in Keras.

• Python 3.6
• Keras 2.2.4

Append CosineAnnealingScheduler to list of callbacks and pass to .fit() or .fit_generator():

from cosine_annealing import CosineAnnealingScheduler

callbacks = [
    CosineAnnealingScheduler(T_max=100, eta_max=1e-2, eta_min=1e-4)
]

model.fit(x, y, batch_size=32, callbacks=callbacks)

Use CosineAnnealingScheduler:

python train.py --scheduler CosineAnnealingScheduler