Keras list of learning rates schedule

list_rate_schedule.py

def list_rate_schedule(lrates, output=True):
    sched = []
    last_lr = [0]

    for lr, n in lrates:
        sched += [lr] * n

    def lr_sched(epoch):
        lr = sched[-1]
        if epoch < len(sched):
            lr = sched[epoch]
        if output and lr != last_lr[0]:
            print('Learning rate: {}'.format(lr))
            last_lr[0] = lr
        return lr

    return lr_sched

# usage ----------------------------------------------------

rates = [
    (1e-5, 2),  # 2 epochs @ 1e-5
    (1e-3, 4),  # 4 epochs @ 1e-3
    (1e-4, 8),  # 8 epochs @ 1e-4
    # ...
]

lrsched = keras.callbacks.LearningRateScheduler(
    list_rate_schedule(rates))

model.fit_generator(
  trn_batches,
  trn_batches.samples // trn_batches.batch_size,
  callacks=[lrsched],
  # ...
)