서론

Huggingface Custom Trainer 작성중에 lr을 정하기 위해 scheduler를 확인 중이었는데 좋은 블로그가 나와서 사용해보았습니다. 그런데 추천하신 스케쥴러가 custom으로 작성하신 코드였는데 Huggingfcace에 적용해보니 에러가 나와서 수정하였습니다.

본론

class CosineAnnealingWarmUpRestarts(_LRScheduler):
    def __init__(self, optimizer, T_0, T_mult=1, eta_max=0.1, T_up=0, gamma=1., last_epoch=-1):
        if T_0 <= 0 or not isinstance(T_0, int):
            raise ValueError("Expected positive integer T_0, but got {}".format(T_0))
        if T_mult < 1 or not isinstance(T_mult, int):
            raise ValueError("Expected integer T_mult >= 1, but got {}".format(T_mult))
        if T_up < 0 or not isinstance(T_up, int):
            raise ValueError("Expected positive integer T_up, but got {}".format(T_up))
        self.T_0 = T_0
        self.T_mult = T_mult
        self.base_eta_max = eta_max
        self.eta_max = eta_max
        self.T_up = T_up
        self.T_i = T_0
        self.gamma = gamma
        self.cycle = 0
        self.T_cur = last_epoch
        super(CosineAnnealingWarmUpRestarts, self).__init__(optimizer, last_epoch)
    
    def get_lr(self):
        if self.T_cur == -1:
            return self.base_lrs
        elif self.T_cur < self.T_up:
            return [(self.eta_max - base_lr)*self.T_cur / self.T_up + base_lr for base_lr in self.base_lrs]
        else:
            return [base_lr + (self.eta_max - base_lr) * (1 + math.cos(math.pi * (self.T_cur-self.T_up) / (self.T_i - self.T_up))) / 2
                    for base_lr in self.base_lrs]

    def step(self, epoch=None):
        if epoch is None:
            epoch = self.last_epoch + 1
            self.T_cur = self.T_cur + 1
            if self.T_cur >= self.T_i:
                self.cycle += 1
                self.T_cur = self.T_cur - self.T_i
                self.T_i = (self.T_i - self.T_up) * self.T_mult + self.T_up
        else:
            if epoch >= self.T_0:
                if self.T_mult == 1:
                    self.T_cur = epoch % self.T_0
                    self.cycle = epoch // self.T_0
                else:
                    n = int(math.log((epoch / self.T_0 * (self.T_mult - 1) + 1), self.T_mult))
                    self.cycle = n
                    self.T_cur = epoch - self.T_0 * (self.T_mult ** n - 1) / (self.T_mult - 1)
                    self.T_i = self.T_0 * self.T_mult ** (n)
            else:
                self.T_i = self.T_0
                self.T_cur = epoch
                
        self.eta_max = self.base_eta_max * (self.gamma**self.cycle)
        self.last_epoch = math.floor(epoch)
        for param_group, lr in zip(self.optimizer.param_groups, self.get_lr()):
            param_group['lr'] = lr

결론

잘 돌아갑니다.

출처

Pytorch Learning Rate Scheduler