深度学习中的LR_Scheduler

写一个例子,以防后面每次用到都忘记

class LR_Scheduler(object):
    def __init__(self, optimizer, warmup_epochs, warmup_lr, num_epochs, base_lr, final_lr, iter_per_epoch,
                 constant_predictor_lr=False):
        self.base_lr = base_lr
        self.constant_predictor_lr = constant_predictor_lr
        warmup_iter = iter_per_epoch * warmup_epochs
        warmup_lr_schedule = np.linspace(warmup_lr, base_lr, warmup_iter)
        decay_iter = iter_per_epoch * (num_epochs - warmup_epochs)
        cosine_lr_schedule = final_lr + 0.5 * (base_lr - final_lr) * (
                    1 + np.cos(np.pi * np.arange(decay_iter) / decay_iter))

        self.lr_schedule = np.concatenate((warmup_lr_schedule, cosine_lr_schedule))
        self.optimizer = optimizer
        self.iter = 0
        self.current_lr = 0

    def step(self):
        for param_group in self.optimizer.param_groups:

            if self.constant_predictor_lr and param_group['name'] == 'predictor':
                param_group['lr'] = self.base_lr
            else:
                lr = param_group['lr'] = self.lr_schedule[self.iter]

        self.iter += 1
        self.current_lr = lr
        return lr

    def get_lr(self):
        return self.current_lr
scheduler = LR_Scheduler(optimizer, warmup_epochs=300, warmup_lr=2e-2, num_epochs=2000, base_lr=3e-2, final_lr=2e-3, iter_per_epoch=1)
lr = []
for i in range(2000):
    lr.append(scheduler.step())
plt.plot(lr)
plt.show()

深度学习中的LR_Scheduler_第1张图片

你可能感兴趣的:(pytorch,python)