AdaX (带”长期记忆“的优化器) _基于pytorch的算法实现

最近，一篇名为《AdaX: Adaptive Gradient Descent with Exponential Long Term Memory》的文章中介绍了一种名为AdaX的优化器。
AdaX优化器在基于Adam优化器的基础上进行了修改。主要是在Adam优化器中二阶梯度的滑动平均$v_t$处做了修改，并且删除了Adam优化器中对于动量 (一阶梯度的滑动平均) $m_t$的偏置校正。
AdaX优化器与Adam优化器的过程对比如下所示：
AdaX优化器的参数更新过程为(其中${\beta }_1$为0.9，${\beta }_2$为0.0001)：

Adam优化器的参数更新过程为(其中${\beta }_1$为0.9，${\beta }_2$为0.999)：

从AdaX优化算法的更新过程中可以看出，AdaX删除了Adam优化器中对于动量 (一阶梯度的滑动平均) $m_t$的偏置校正；并且AdaX优化算法在$v_t$处的形式为:

此时(1 + ${\beta }_2$) > 1，导致历史累积梯度的比重不会越来越小，反而会越来越大，这就是它的长期记忆性，也即为原论文中所提到的“with Exponential Long Term Memory”。

原论文中AdaX优化算法的伪代码如下所示：

原论文的链接为：AdaX: Adaptive Gradient Descent with Exponential Long Term Memory
另一篇更全面的关于AdaX优化算法的解析可以参考 夕小瑶的卖萌屋的文章：AdaX：一个比Adam更优秀，带”长期记忆“的优化器

AdaX优化算法的pytorch实现如下所示：

import math
import torch
# from .optimizer import Optimizer
from torch.optim import Optimizer


class AdaX(Optimizer):
    r"""Implements AdaX algorithm.

    Arguments:
        params (iterable): iterable of parameters to optimize or dicts defining
            parameter groups
        lr (float, optional): learning rate (default: 1e-3)
        betas (Tuple[float, float], optional): coefficients used for computing
            running averages of gradient and its square (default: (0.9, 0.999))
        eps (float, optional): term added to the denominator to improve
            numerical stability (default: 1e-8)
        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
        amsgrad (boolean, optional): whether to use the AMSGrad variant of this
            algorithm from the paper `On the Convergence of Adam and Beyond`_
            (default: False)

    .. _Adam\: A Method for Stochastic Optimization:
        https://arxiv.org/abs/1412.6980
    .. _On the Convergence of Adam and Beyond:
        https://openreview.net/forum?id=ryQu7f-RZ
    """

    def __init__(self, params, lr=1e-3, betas=(0.9, 0.0001), eps=1e-8,
                 weight_decay=0, amsgrad=False):
        if not 0.0 <= lr:
            raise ValueError("Invalid learning rate: {}".format(lr))
        if not 0.0 <= eps:
            raise ValueError("Invalid epsilon value: {}".format(eps))
        if not 0.0 <= betas[0] < 1.0:
            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
        if not 0.0 <= betas[1] < 1.0:
            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
        if not 0.0 <= weight_decay:
            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
        defaults = dict(lr=lr, betas=betas, eps=eps,
                        weight_decay=weight_decay, amsgrad=amsgrad)
        super(AdaX, self).__init__(params, defaults)

    def __setstate__(self, state):
        super(AdaX, self).__setstate__(state)
        for group in self.param_groups:
            group.setdefault('amsgrad', False)

# [docs]    @torch.no_grad()
    def step(self, closure=None):
        """Performs a single optimization step.

        Arguments:
            closure (callable, optional): A closure that reevaluates the model
                and returns the loss.
        """
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            for p in group['params']:
                if p.grad is None:
                    continue
                grad = p.grad
                if grad.is_sparse:
                    raise RuntimeError('AdaX does not support sparse gradients, please consider SparseAdam instead')
                amsgrad = group['amsgrad']

                state = self.state[p]

                # State initialization
                if len(state) == 0:
                    state['step'] = 0
                    # Exponential moving average of gradient values
                    state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format)
                    # Exponential moving average of squared gradient values
                    state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
                    if amsgrad:
                        # Maintains max of all exp. moving avg. of sq. grad. values
                        state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)

                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
                if amsgrad:
                    max_exp_avg_sq = state['max_exp_avg_sq']
                beta1, beta2 = group['betas']
                
                # state['step']表示迭代次数t.
                state['step'] += 1
                # 并且此处删除了Adam优化器中对于动量 (一阶梯度的滑动平均) $m_{t}$的偏置校正bias_correction1.
                # 此处(1 + beta2) > 1导致历史累积梯度的比重不会越来越小，反而会越来越大，这就是它的长期记忆性,
                # 也即为原论文中所提到的“with Exponential Long Term Memory”.
                bias_correction2 = (1 + beta2) ** state['step'] - 1  # modified

                if group['weight_decay'] != 0:
                    grad = grad.add(p, alpha=group['weight_decay'])

                # Decay the first and second moment running average coefficient
                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
                exp_avg_sq.mul_(1 + beta2).addcmul_(grad, grad, value=beta2)  # modified
                
                if amsgrad:
                    # Maintains the maximum of all 2nd moment running avg. till now
                    torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)
                    # Use the max. for normalizing running avg. of gradient
                    denom = (max_exp_avg_sq.sqrt() / math.sqrt(bias_correction2)).add_(group['eps'])
                else:
                    denom = (exp_avg_sq.sqrt() / math.sqrt(bias_correction2)).add_(group['eps'])
                
                # modified
                step_size = group['lr']

                p.addcdiv_(exp_avg, denom, value=-step_size)

        return loss

References
[1] Li W, Zhang Z, Wang X, et al. AdaX: Adaptive Gradient Descent with Exponential Long Term Memory[J]. arXiv preprint arXiv:2004.09740, 2020.